University of Southern California Dissertations and Theses

Neuromorphic motion sensing circuits in a silicon retina

(USC Thesis Other)

Neuromorphic motion sensing circuits in a silicon retina

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content NEUROMORPHIC MOTION SENSING CIRCUITS IN A SILICON RETINA
by
Ko-Chung Tseng
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)
December 2012
Copyright 2012 Ko-Chung Tseng
I lovingly dedicate this thesis to my parents and my wife, who
supported me each step of the way.
ii
Acknowledgments
Upon the completion of my Ph.D thesis, I would like to acknowledge the people who
have contributed to the preparation and completion of this study. I feel that my achieve-
ments are due directly to the support and guidance of a few key players.
Dr. Parker has been my advisor over the course of my study. She has had a direct
influence on my success as a student and individual. Her patience and guidance has
allowed me to grow as a scholar and professional. Thank you Dr. Parker, your presence
as a mentor and coach has resulted in my curiosity and pursuit of excellence in life. My
experience with you has been very pleasurable and will always reside in my memory as
some of the best times in my life.
Thank you mom and dad, without you nothing would have been possible. I love you
both dearly and will repay your support with becoming a valued member of society. To
my wife, thank you for looking over me and taking care of me. You are my soul mate
and your companionship during my study means the world to me.
I still remember receiving my acceptance letter from this prestigious university. The
feelings of excitement and gratitude still reside in my memory to this day. I would like to
acknowledge and thank this great university for giving me the opportunity and training
to understand the importance of personal growth and the pursuit of greater knowledge.
I will remain forever grateful and will, one day, share my prosperity and success with
the future students and educators of this great facility.
iii
Table of Contents
Dedication ii
Acknowledgments iii
List of Tables vi
List of Figures vii
2.3 Comparison of the Research to State of the Art in Neuromorphic Designs
of the Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
iv
1.2 Hypothesis Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Chapter 2 Background and Related Work 6
2.1.1 The Outer Retina . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 The Inner Retina . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Related Retinal Neuromorphic Research . . . . . . . . . . . . . . . . . 16
2.2.1 Neuromorphic Designs of the Retina in Silicon Circuits . . . . . 16
2.2.2 Neuromorphic Models of Motion Detection in the Visual System 19
2.2.3 Implantable Artificial Retinas . . . . . . . . . . . . . . . . . . 20
Chapter 3 The Outer Retina Design 24
3.1 The Photoreceptor Circuit and Testing Results . . . . . . . . . . . . . . 24
3.2 Horizontal Cell Design . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 The Outer Retina Network and Testing Results . . . . . . . . . . . . . 34
3.3.1 Glutamate Reuptake . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.2 Contrast Enhancement . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Center-Surround Property . . . . . . . . . . . . . . . . . . . . 42
2.1 Biological Background . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Abstract xi
Chapter 1 Introduction 1
3.3.4 Edge Detection . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Synapses in Starburst Amacrine Cells . . . . . . . . . . . . . . . . . . 52
4.3 A Neuromorphic Circuit that Computes Differential Motion . . . . . . 66
4.4 Retinal Pathways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4.1 Object-Motion-Sensitivity (OMS) Ganglion Cell Pathway . . . 88
4.4.2 On-Center Ganglion Cell Pathway . . . . . . . . . . . . . . . . 93
4.4.3 Directionally-Selective Ganglion Cell Pathway . . . . . . . . . 97
References 105
Appendix 112
v
Chapter 4 The Inner Retina Design 48
4.1 On-type Bipolar Cell Design . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 A Directionally-Selective Neuromorphic Circuit Based on Reciprocal
Chapter 5 Conclusion and Future Research 101
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
List of Tables
2.1 Comparison to State of the Art Research in Neuromorphic Designs of
the Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
vi
List of Figures
2.1 A Cross-Sectional View of an Eye . . . . . . . . . . . . . . . . . . . . 6
2.2 A Cross-Sectional View of a Biological Retina[17] . . . . . . . . . . . 8
2.3 A Retinal Circuitry Showing Major Pathways . . . . . . . . . . . . . . 9
2.4 Functional Diagram of the Interaction Between Cone and Horizontal Cell 11
2.5 The Reciprocal Interactions Between Two Branches of the SACs . . . . 14
2.6 The Interaction of a Direction-Selective Ganglion Cell (DSGC) with
Bipolar Cells and SACs . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 The Underlying Retinal Circuitry for Performing Differential Motion
Detection and Saccadic Suppression . . . . . . . . . . . . . . . . . . . 17
3.1 The Photoreceptor Circuit Design By Delbruck and Mead . . . . . . . . 25
3.2 The Photoreceptor Circuit Design . . . . . . . . . . . . . . . . . . . . 26
3.3 Simulation Result of an Isolated Photoreceptor . . . . . . . . . . . . . 27
3.4 Response Ranges of the Photoreceptor (When Sensitivity is 900 mV) . . 28
3.5 The Range of the Responses of an Isolated Photoreceptor Given Differ-
ent Sensitivity Values Ranging from 0.6 to 0.3 V olt . . . . . . . . . . . 29
3.6 Best-Fit Line Representing the Response of the Photoreceptor Circuit . 30
3.7 Curves Representing the Decomposed Terms of the Best-Fit Line . . . . 31
3.8 Comparison of the photoreceptor Output to Biological Photoreceptors . 32
vii
3.9 The Horizontal Cell Compartment Design . . . . . . . . . . . . . . . . 33
3.10 The Interaction Between a Photoreceptor Design and a Horizontal Cell
Compartment Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.11 A One-Dimensional Photoreceptor-Horizontal Cell Network . . . . . . 36
3.12 Simulation Result for Changing Glutamate Reuptake from -0.9 to 0.9 V olt 38
3.13 Two Configurations for Testing Contrast Enhancement . . . . . . . . . 40
3.14 Simulation Results of Contrast Enhancement . . . . . . . . . . . . . . 41
3.15 The Configuration of the 14 by 14 Photoreceptor-Horizontal Cell (PHC)
Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.16 Demonstration of the Center Surround Property . . . . . . . . . . . . . 45
3.17 The Responses of Photoreceptors to an Edge Over Time. . . . . . . . . 47
4.1 The Post-Synaptic Circuit of the Transient-ON Bipolar Cell . . . . . . . 50
4.2 Simulation Result of the Post-Synaptic Circuit of the Transient-ON Bipo-
lar Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Simplified Transient-ON Bipolar Cell Model with Dendritic Computation 51
4.4 The Inhibitory Interaction Between Two SACs . . . . . . . . . . . . . . 54
4.5 Starburst Amacrine Cell . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.6 Basic Distal Compartment of the SAC . . . . . . . . . . . . . . . . . . 56
4.7 Two Overlapped Branches of Two Simplified Biological Starburst Amacrine
Cells (SACs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.8 Simulation Results of a Single SAC with Respect to Both Centripetal
and Centrifugal Motion . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.9 Comparison of the Simulation Results Both with and without Reciprocal
Synapse to Centrifugal Motion . . . . . . . . . . . . . . . . . . . . . . 62
4.10 Amplitude of the SAC Distal Compartment vs. Speed of the stimulus . 63
4.11 Amplitude of the SAC Distal Compartment vs. Input Intensity . . . . . 64
viii
4.12 DS Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.13 A Neuromorphic Circuit that Performs Differential Motion . . . . . . . 68
4.14 A Complete Bipolar Cell Circuit Model . . . . . . . . . . . . . . . . . 68
4.15 Simulation Results of Differential Motion Detection Circuit, Showing
the Summation of the Bipolar Cells’ Responses over Time as a grating
is shifted . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.16 Maximum Bipolar Cell Responses under Different Speed Combinations 72
4.17 Responses over Different Input Intensities to Moving Gratings . . . . . 73
4.18 Responses in Bar Chart I: The Bar Spacing is 5 and Vary the Bar Width. 74
4.19 Responses in Bar Chart II: The Bar Width is 1 and Vary the Bar Spacing. 75
4.20 Responses in Bar Chart III: The Bar Width is 2 and Vary the Bar Spacing. 76
4.21 Responses in Bar Chart IV: The Bar Width is 3 and Vary the Bar Spacing. 77
4.22 Comparing Responses in Curves I: The Bar Spacing is 5 and Vary the
Bar Width. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.23 Comparing Responses in Curves II: The Bar Width is 1 and Vary the
Bar Spacing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.24 Comparing Responses in Curves III: The Bar Width is 2 and Vary the
Bar Spacing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.25 Comparing Responses in Curves IV: The Bar Width is 3 and Vary the
Bar Spacing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.26 Enhancement I: The Bar Spacing is 1 and Vary the Bar Width. . . . . . 84
4.27 Enhancement II: The Bar Width is 1 and Vary the Bar Spacing. . . . . . 85
4.28 Enhancement III: The Bar Width is 2 and Vary the Bar Spacing. . . . . 85
4.29 Enhancement IV: The Bar Width is 3 and Vary the Bar Spacing. . . . . 86
4.30 Spiking Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.31 Measure the Outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
ix
4.32 Responses at the Same Speed . . . . . . . . . . . . . . . . . . . . . . . 90
4.33 Responses at the Different Speeds . . . . . . . . . . . . . . . . . . . . 91
4.34 Comparison to the Biological Results . . . . . . . . . . . . . . . . . . 92
4.35 On-Center Ganglion Cell Pathway . . . . . . . . . . . . . . . . . . . . 94
4.36 Simulation Result of On-Center Ganglion Cell Pathway I . . . . . . . . 95
4.37 Simulation Result of On-Center Ganglion Cell Pathway II . . . . . . . 96
4.38 On Center Bipolar Cell Circuit Model . . . . . . . . . . . . . . . . . . 96
4.39 Directionally-Selective Ganglion Cell Pathway . . . . . . . . . . . . . 97
4.40 Directionally-Selective Ganglion Cell (DSGC) Circuit model . . . . . . 98
4.41 The Response of DSGC to Centrifugal Motion . . . . . . . . . . . . . . 99
4.42 The Response of DSGC to Centripetal Motion . . . . . . . . . . . . . . 100
5.1 CMOS Photoreceptor Circuit . . . . . . . . . . . . . . . . . . . . . . . 113
5.2 Glutamate Reuptake and Horizontal Cell Compartment Circuit . . . . . 114
5.3 CMOS Bipolar Cell Circuit . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4 CMOS Starburst Amacrine Cell Circuit . . . . . . . . . . . . . . . . . 116
5.5 CMOS Wide Field Amacrine Cell Circuit . . . . . . . . . . . . . . . . 117
5.6 The Sublinear V oltage Adder in Object-Motion-Sensitive Ganglion Cell 118
5.7 The Single Spiking Circuit . . . . . . . . . . . . . . . . . . . . . . . . 119
5.8 The Directionally-Selective Ganglion Cell Circuit . . . . . . . . . . . . 120
5.9 On Bipolar Cell and On Ganglion Cell for Demonstrating Center-Surround
Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
x
Abstract
In the biological retina, the feedback and lateral pathways among retinal neurons con-
struct a complicated network that contributes to motion sensing in the retina. When
these complex pathways and diverse retinal cell types collaborate, the retina effectively
extracts useful information from the visual scene and communicates it to the brain. A
silicon retina with motion sensing may be useful for service robots, autonomous vehi-
cles and other applications that require processing dynamic visual information in real
time. Implementing motion sensing in a silicon retina presents many challenges. For
engineers trying to model the motion sensing functions of a silicon retina, connectivity
is one of the most significant engineering challenges that have to be considered.
For this dissertation research, we implement a portion of the starburst amacrine cell
(SAC) and differential motion detection model. We also investigate the importance of
the feedback and lateral connections in implementing these motion sensing functions
in silicon circuit. To validate the importance of the feedback and lateral pathways in
the silicon retina, we first build a portion of a retinal network from photoreceptors to
ganglion cells that maintains a hierarchical structure similar to that of the biological
retina. Lateral connections with horizontal cells and amacrine cells are implemented,
along with feedback within the inner and outer plexiform layers of the retina. We then
perform demonstrations by comparing the silicon retina tested to one altered by remov-
ing these pathways and observing how the behaviors in the silicon retina are changed.
xi
We also compare some of our simulation results with biological data. In this research,
we showed that some functions cannot be achieved or performances degrade without
feedback and lateral connections. Hence, we concluded that incorporating feedback and
lateral connections in the artificial retina helps the performance even though it compli-
cates the retinal network.
xii
Chapter 1
Introduction
1.1 Motivation
The retina preprocesses visual information with millions of neurons and glial cells, and
with billions of synapses. Motion sensing in particular is found in some vertebrate reti-
nas. A silicon retina with motion sensing may be useful for service robots, autonomous
vehicles and other applications that require processing dynamic visual information in
real time. Implementing motion sensing in a silicon retina presents many challenges.
In the biological retina, the feedback and lateral pathways among retinal neurons con-
stitute a complicated network that contributes to motion sensing in the retina. When
these complex pathways and diverse retinal cell types collaborate, the retina effectively
extracts useful information from the visual scene and communicates it to the brain. For
engineers trying to model the motion-sensing functions of a silicon retina, connectivity
is one of the most significant engineering challenges that has to be considered.
Including the feedback and lateral pathways in the silicon retina makes modeling a
retinal network challenging. Firstly, the feedback pathways in the silicon retina could
potentially lead to an unstable system. The propagation delays and the signal strength
over lateral pathways need to be properly tuned. Bi-directional signaling over the lat-
eral pathways may form self-feedback loops that could also lead to an unstable system.
Hence, engineers must ensure the stability of the system’s operation while modeling the
functions of a biological retina.
1
There are at least 50 distinct cell types that have been found. The retinal neurons
are arranged in three cellular layers (i.e. outer nuclear layer, inner nuclear layer, and
retinal ganglion cell layer) and are interconnected in the intervening two synaptic lay-
ers (i.e. outer plexiform layer and inner plexiform layer). The outer plexiform layer
is a layer of neuronal synapses that consists of a dense network of synapses between
dendrites of horizontal cells, bipolar cells (from the inner nuclear layer), and photore-
ceptor cell inner segments (from the outer nuclear layer); The inner plexiform layer is a
layer of neural synapses consisting of a dense reticulum of fibrils formed by interlaced
dendrites of retinal ganglion cells, bipolar cells, and amacrine cells. Lateral communi-
cations formed by the amacrine cells and horizontal cells are noteworthy because they
increase connectivity significantly. The feedback and lateral communications found in
these two layers are believed to be important to modulate the responses of the retinal
network in order to perform some essential functions. For instance, the existence of
lateral connections helps to employ center-surround antagonism in the ganglion cell’s
receptive field that may sharpen the image in time and space. Neuroscientists have found
another example in which starburst amacrine cells (SACs) in the vertebrate retina (such
as the rabbit’s retina) perform the neural computations that induce directional selectivity
in the ganglion cell. Therefore, the ganglion cell performs directional selectivity. Other
functions such as differential motion detection have also been observed in the vertebrate
retina[25]. To perform these functions, the feedback and lateral signaling pathways play
important roles.
For this dissertation research, we implement a portion of the starburst amacrine cell
(SAC) and differential motion detection model. We also investigate the importance of
the feedback and lateral connections in implementing these motion sensing functions in
silicon circuit. And, raise several questions. Are modeling feedback and lateral connec-
tions required to perform the observed motion sensing functions or behaviors found in
2
biological retinas? How does the response (or behavior) change after removing those
connections from the silicon retina? Having feedback and lateral connections compli-
cates the silicon retina. Does incorporating those connections perform a useful function?
1.2 Hypothesis Statement
Feedback and lateral interactions are important in a silicon retina to model the response
of a vertebrate retina, including saccadic suppression (differential motion detection) and
directional selectivity.
To effectively compute specific features of the visual input from the visual scene, the
feedback and lateral pathways in the biological retina allow neurons to communicate
with each other and mediate the neuronal responses. Therefore, in this research we
hypothesize that the silicon retina needs to incorporate the mechanisms of feedback and
lateral interactions to achieve certain observed behaviors or to enhance performance. If
we remove these mechanisms, the behavior may fail to perform the observed behavior or
performances degrade. However, including feedback and lateral interactions in a silicon
retina complicates the connectivity. In this dissertation, we especially focus on motion
sensing functions in the retina.
Apart from this, modeling motion sensing in the retina presents many challenges:
Biological retinal networks are complicated with feedback and lateral pathways.
Feedback pathways could potentially lead to an unstable system. Bi-directional
signaling over lateral pathways may form self-feedback loops which may also
cause an unstable system. The propagation delays over the pathway need to be
properly tuned.
3
Current CMOS technology provides a limited amount of die area and number of
metal layers for interconnections. Hence, we may need advanced technology to
complete the silicon retina.
1.3 Thesis Outline
Chapter 2 starts by providing a review of biological background that covers basic struc-
ture of the biological retina (i.e. outer retina and inner retina), the computation models
of directional selectivity involving the starburst amacrine cell (SAC), and differential
motion detection model (a.k.a. saccadic suppression model). Moreover, we compare
the related retinal neuromorphic research with our work in this chapter.
Chapter 3 presents the outer retinal circuits that we have built as well as their test-
ing results. The outer retina has been modeled by others extensively. In order to be
compatible with the inner retinal circuits in Chapter 4, we built our own outer retinal
circuits. The circuit models include the photoreceptor circuit, the horizontal cell com-
partment circuits, and the transient on-type bipolar cell circuit. Contrast enhancement,
edge detection, and center surround antagonism are demonstrated using the outer retinal
circuits in this chapter.
Chapter 4 presents the inner retinal circuits that we have built as well as their testing
results. The research aims at modeling the first-order responses (or behaviors) of the
intermediate retinal cells by considering their underlying (both inter-cellular and intra-
cellular) mechanisms. We present a compartmentalized CMOS neuromorphic circuit
that models a portion of two biological starburst amacrine cells in the retina and includes
a simplified model of reciprocal interaction between the dendritic branches of SACs. We
demonstrate that a neuromorphic circuit incorporating the reciprocal synapses enhances
4
the responses in the neuromorphic dendritic tip and generates robust directional selec-
tivity. In chapter 4, we also present a neuromorphic circuit that compares the motion
speeds of the central receptive field and peripheral receptive field. We demonstrate that
the response is suppressed if motion speeds of the central receptive field and the periph-
eral receptive field are the same. To validate the importance of the feedback and lateral
pathways in the silicon retina, we performed demonstrations by comparing the silicon
retina tested to one altered by removing these pathways and observing how the behaviors
in the silicon retina are changed. We also performed analysis showing how much the per-
formance was improved relative to hardware cost. Finally, we have built a portion of a
retinal network in silicon CMOS technology, from photoreceptors to ganglion cells, that
maintains a hierarchical structure similar to that of the biological retina and observed
the spiking behaviors. The lateral connections with horizontal cells and amacrine cells
are implemented, along with feedback within the inner and outer plexiform layers of the
retina.
Finally, conclusions have been drawn in Chapter 5 and future research topics have
been suggested and further discussed.
5
Chapter 2
Background and Related Work
2.1 Biological Background
The retina is a complex multi-layered neural tissue consisting of many neurons inter-
connected by synapses and located in the back of the eye. In mammalian retinas, the
output is conveyed to the brain by about 15 different retinal ganglion cell (RGC) types.
Figure 2.1 shows a simple anatomy of the eye.
Light is first absorbed by photoreceptors and then converted into electrical signals,
namely graded potentials. There are two main types of photoreceptors in the retina: rods
and cones. Compared to the rod cells, the cone cells are less sensitive to light, but allow
Figure 2.1: A Cross-Sectional View of an Eye
6
the perception of color. Furthermore, the cone cells become sparser towards the periph-
ery of the retina while the rod cells become denser towards the periphery of the retina.
Horizontal cells that synapse with the photoreceptor cells can facilitate the communica-
tion between photoreceptors. Horizontal cells are believed to play a key role in forming
center-surround antagonism, which enables edge detection and contrast enhancement.
The signals from photoreceptor cells and horizontal cells are relayed by bipolar cells
through which they reach amacrine cells and ganglion cells. There are about 40 differ-
ent types of amacrine cells in the retina. They work laterally affecting other amacrine
cells, bipolar cells, and ganglion cells. It is widely accepted that amacrine cells perform
more specialized tasks than horizontal cells. The retinal signals from bipolar cells and
amacrine cells reach ganglion cells which convert graded potentials into action poten-
tials, namely spikes. Some neuroscientists believe that each ganglion cell type computes
rather differently, based on the contents of the visual scene [25]. Lastly, the spikes from
retinal ganglion axons forming the optic nerve are delivered to the visual cortex. Besides
these retinal cells, retinal glial cells spanning the entire thickness of the retinal layers
may modulate the electrical activity of neurons within the retina[65][29]. There are three
basic types of glial cells found in the human retina: M¨ uller, astroglia, and microglia cells
[42]. Conventionally, M¨ uller cells are the principal glial cell of the retina and believed to
be vital to the health of the retinal neurons [42] [8]. A two-way communication between
retinal neurons and glial cells may exist and suggest that the glia contributes to infor-
mation processing in the retina[55]. For the basic structure of a retina, please refer to
Figure 2.2 that shows the organization of the principal retinal cells.
Figure 2.3 illustrates the major pathways of principle retinal cells. As described
previously, photoreceptors and horizontal cells form reciprocal synapses that help to
increase the dynamic range of photoreceptors and to perform many functions such as
7
Figure 2.2: A Cross-Sectional View of a Biological Retina[17]
contrast enhancement and center-surround antagonism. The channels between hori-
zontal cells are confirmed to be gap junctions that facilitate the communication across
different horizontal cells. The pathway from horizontal cell to bipolar cell may be weak
[78]. Bipolar cells relay the signal from photoreceptors to amacrine cells, as well as
to retinal ganglion cells. They may also receive feedback from amacrine cells. Sim-
ilar to horizontal cells, reciprocal synapses between amacrine cells are also identified.
They may be electrical synapses or chemical synapses. Ganglion cells may synapse
with amacrine cells as well as bipolar cells, and then convert the graded potentials into
action potentials, namely spikes. M¨ uller cells can influence the response of retinal cells
[56]. Note that Figure 2.3 only illustrates the principal neurons of the retina that we will
discuss in the following sections.
8
Figure 2.3: A Retinal Circuitry Showing Major Pathways
2.1.1 The Outer Retina
In the outer retina, the photoreceptors absorb photons and convert them into electrical
signals that modulate the release of glutamate from photoreceptor ribbon synapses. The
horizontal cells (HCs) receive the glutamate released from photoreceptors and provide
feedback to the photoreceptors. The feedback from the horizontal cells to the photore-
ceptor cells helps to ensure the transmission of sensory patterns with proper contrast,
modulates the dynamic range of retinal cells’ responses, and enhances spatial contrast
in the retina[71]. Three synaptic mechanisms have been proposed to explain the under-
lying mechanisms for the feedback[72]. The first mechanism is that HCs release the
9
inhibitory GABA neurotransmitter in darkness, which opens chloride channels in cones.
When a light stimuli hyperpolarizes the HCs, HCs release less GABA neurotransmitter,
which suppresses feedback transmitter release and depolarizes the cones[52][70]. The
second theory is that a light stimulus hyperpolarizes HCs, resulting in an inward current
through hemichannels in their dendrites near the cones, charging the cone membrane
and modulating calcium currents in cones, increasing their calcium-dependent gluta-
mate release [39] [38]. The third theory is that light-induced HC hyperpolarization
elevates the pH in the HC-cone synaptic cleft, leading to an increase in calcium current
in the cones[68][30].
The horizontal cells may communicate with their neighboring horizontal cells
through gap junctions whose permabilities are mediated by some neurotransmitters [74].
Both the photoreceptors and horizontal cells have connections with the bipolar cells.
However, neuroscientists found that the strength of the connections from horizontal
cells to bipolar cells may be weak [78]. Thus, the feedback from the horizontal cells
to the photoreceptors that regulates the responses of photoreceptors is believed to play
the more important role in producing some important computations observed in bipolar
cells. Due to the lateral communications between the horizontal cells and the feedback
from the horizontal cells to photoreceptor cells, the outer retina is, therefore, considered
the start of the center-surround mechanism in the whole visual system. Eventually, the
visual information is relayed by the bipolar cells and sent to the inner retina.
We show the interaction between the photoreceptors and horizontal cells in Figure
2.4. In a biological retina, one horizontal cell can synapse with multiple photoreceptors
in the retina. But, we only show one photoreceptor, which is a cone, in Figure 2.4. When
the photon enters the retina, it is first absorbed by the photoreceptor. The photoreceptor
converts the light into electrical signals, i.e. the phototransduction process. During the
resting state, the photoreceptor releases glutamate tonically. Hyperpolarization of the
10
Figure 2.4: Functional Diagram of the Interaction Between Cone and Horizontal Cell
photoreceptor will reduce the release of glutamate, which is an excitatory neurotrans-
mitter. Since the horizontal cell receives less excitation from the photoreceptors through
ribbon synapses, the horizontal cell, therefore, hyperpolarizes as well. Hyperpolariza-
tion of the horizontal cell provides an inhibition to the photoreceptor as a net effect. The
overall interaction is analogous to provide a negative feedback from the horizontal cell
to the photoreceptors.
A positive feedback synapse from horizontal cells to cone photoreceptors has
recently been found in the outer retina[34]. Unlike the negative feedback that spreads
throughout a horizontal cell to affect many surrounding photoreceptors, the positive
feedback signal is constrained to individual horizontal cell-photoreceptor connections.
The positive feedback locally offsets the effects of negative feedback and amplifies pho-
toreceptor synaptic release without sacrificing HC-mediated contrast enhancement.
11
2.1.2 The Inner Retina
The inner retina is believed to be the place where retinal motion detection originates.
The bipolar cell relays the visual information from the outer retina and sends it to the
amacrine cells and ganglion cells. In the inner retina, multiple feedback pathways have
been identified recently between amacrine cells [32] and between amacrine cells and
bipolar cells [18][48]. With these feedback pathways and the collaboration of diverse
retinal cell types, the inner retina may perform various retinal computations such as dif-
ferential motion detection[4] [58], directional selectivity [50] [43], approaching motion
detection [51], and others. In addition to the feedback, lateral communication also con-
tributes to regulating the responses of other retinal cells [11]. We will describe two
functions in this section: directional selectivity and differential motion detection.
Directional Selectivity
The starburst amacrine cell (SAC), found in the mammalian retina, with a character-
istic radially symmetric morphology, is thought to provide directional inhibitory input
to direction-selective ganglion cells (DSGCs)[76][1][22]. The computation of direc-
tion selectivity (DS) occurs at individual dendritic branches of each SAC and each
dendritic branch acts as an independent computation module[21]. Both the dendritic
calcium signal and membrane voltage in the dendritic tip may generate a stronger
response by the stimuli moving from the soma towards the dendritic tip (namely cen-
trifugal motion) than moving the opposite direction (namely centripetal motion)[21].
To explain the DS observed in the SACs, neuroscientists have proposed at least two
fundamentally different mechanisms[28]: dendrite-intrinsic electrotonics[60][66] and
lateral inhibition[7][43]. Euler et al. demonstrated that the intrinsic electrical mecha-
nisms of SACs may produce DS without inhibitory network interactions[28]. Further,
the lateral inhibition between two SACs may enhance the difference in response and
12
generate a robust directional selectivity[43]. SACs may receive glutamate release from
bipolar cells (BCs). Furthermore, the dendritic tip may release and receive GABA neu-
rotransmitter. Euler et al. observed a weak DS at the soma but a strong DS in the
dendritic tips[21]. They also found that SACs may have directional responses even if
GABA inhibitory interactions between the SACs are blocked pharmacologically[21].
Their results suggest DS in the starburst cell arises intrinsically from its distinctive
morphology. A centrifugal (CF) motion generates an in-phase response which can be
summed effectively with the response in the distal compartment. However, centripetal
(CP) motion generates an out-of-phase response that cannot be summed effectively with
the response in the distal compartment. Hence, a SAC produces a stronger response in
the distal tip with respect to centrifugal motion. The lateral inhibition between two over-
lapping SACs may make the DS response more robust[43]. The distal dendrite of SACs
may release GABA neurotransmitters and GABA receptors are also found in the SAC
dendrites. Therefore, as long as the processes of two neighboring starburst cells over-
lap, they are likely to form reciprocal connections. The reciprocal synapse that forms a
positive feedback loop can enhance the difference in response between the two SACs.
The interactions between SACs are illustrated in Figure 2.5. When the light moves to
BC2, the distal dendrite of SAC1 produces a voltage response and releases more GABA,
which inhibits the response of distal dendrites in SAC2. SAC2 in turn produces less
GABA release, which enhances the response of the distal dendrite in SAC1.
Although SACs play an important role in producing DS, the actual DS output in the
retina is at DSGCs. Signaling between the bipolar cell, SAC, and DSGC constitutes
a neural network that generates the DS light responses of the DSGC [19] as show in
Figure 2.6. The glutamate released from the BC is an excitatory neurotransmitter to
the DSGC while the GABA released from SAC is an inhibitory neurotransmitter to the
DSGC. Due to a strong response of SAC1 that induces more GABA release, the DSGC
13
Figure 2.5: The Reciprocal Interactions Between Two Branches of the SACs
is being more strongly inhibited from firing. In the example shown in Figure 2.6, the
DSGC synapses with BC2 and SAC1. The light stimulus moving from BC1 to BC2
evokes a larger response at the distal tip of SAC1 that cancels out the excitatory input
from BC2 and inhibits the DSGC from firing. If the light stimulus moves from BC2 to
BC1, SAC1 does not provide enough inhibitory input to DSGC. Hence, the DSGC fires.
Saccadic Suppression/Object Motion Detection
Large retinal shifts such as saccadic eye movements of humans and other animals may
cause great excitation across the entire retina. These eye movements can be surprisingly
large, up to 0.5

in amplitude (the width of the full moon) and up to 1

per second.
However, humans do not perceptually report this effect. Although the brain-motor com-
mand may suppress the perception of saccadic eye movements, research results suggest
14
Figure 2.6: The Interaction of a Direction-Selective Ganglion Cell (DSGC) with Bipolar
Cells and SACs
that retinal motion alone is sufficient to produce this perceptual suppression [45]. Cer-
tain retinal ganglion cells are strongly inhibited with brief, well-timed inhibition during
saccades [62]. The mechanisms that suppress the visual effects of eye movements are
found in the inner retina.
An object-motion-sensitivity ganglion cell remains silent under global motion of the
entire image but fires when the image patch in its receptive field moves differently from
the background[58][4]. Object-motion-sensitivity ganglion cells meet two seemingly
conflicting requirements: they are highly tuned to a condition of differential motion
between the receptive field center and the surround, but at the same time remarkably
insensitive to the actual visual pattern in the center or the surround. The polyaxonal
amacrine cell appears to be a plausible candidate to transmit inhibition from the back-
ground region[4]. The inhibition signal may derive from amacrine cells that inhibit the
15
bipolar cell synaptic terminal, close to the site of transmission but at some electrotonic
distance from the soma. Only the polyaxonal amacrine cells have the response properties
required to implement inhibition from global motion. The underlying retinal circuitry
for performing differential motion detection is shown in Figure 2.7. The visual inputs
from the peripheral and central receptive fields are relayed by the bipolar cells(BC).
The amacrine cell(AC) delivers timed inhibition from the peripheral receptive field to
suppress the responses of central bipolar cell terminals. The object-motion-sensitivity
ganglion cell (OMS) synapses with the bipolar cells of the central receptive field. An
object-motion-sensitivity (OMS) ganglion cell remains silent under global motion of
the entire receptive field but fires when the image patch in the peripheral receptive field
moves differently from the central receptive field. Moreover, the underlying retinal cir-
cuitry compares only the speed of motion of the object and the background, not the
direction. Both object motion sensitivity and saccadic suppression involving timed inhi-
bition from globally correlated shifting stimuli may share common circuitry.
2.2 Related Retinal Neuromorphic Research
In an effort to compare our research with other related work, we introduce related work
by classifying the related work into three categories: (1) neuromorphic designs of the
retina in silicon circuits, (2) neuromorphic models of motion detection in the visual
system, and (3) implantable artificial retinas. We describe the related studies and explain
the importance of our research by presenting a comparison.
2.2.1 Neuromorphic Designs of the Retina in Silicon Circuits
Mead and Mahowald first modeled early visual processing in the retina[47] in analog
complementary metal-oxide-semiconductor (CMOS) using very large scale integration
16
Figure 2.7: The Underlying Retinal Circuitry for Performing Differential Motion Detec-
tion and Saccadic Suppression
(VLSI) technology. The computation performed by their silicon retina is based on mod-
els of computation in distal layers of the vertebrate retina, which include the cones, the
horizontal cells, and the bipolar cells. Cones have been implemented using parasitic
phototransistors and MOS-diode logarithmic current to voltage converters. Horizontal
cells perform averaging using a hexagonal network of resistors. Bipolar cells detect
the difference between the average output of the horizontal cells and the output of the
cone. The design can perform contrast enhancement and center-surround antagonism.
Soon after that, Delbruck and Mead proposed the first adaptive photoreceptor[12] which
models light adaptation using feedback in an analog circuit. Mahowald combined these
two designs and proposed a more complete artificial outer retina[46] that can also per-
form contrast enhancement and center-surround antagonism. Delbruck later proposed
several improved versions of the adaptive photoreceptor designs[13][14]. Boahen and
17
Andreou modeled the outer plexiform layer of the vertebrate retina using an analog
circuit[6]. They used a current-mode circuit to model the reciprocal synapses between
cones and horizontal cells that produce the antagonistic center/surround receptive field.
Delbruck and Liu designed a silicon chip that emulates the neurons in the visual sys-
tem by using analog Very Large Scale Integration (aVLSI) circuits[15]. Their design
aimed at substituting for a live animal in experiment designs and lectures. The model
contained photoreceptor cells, horizontal cells, On-OFF bipolar cells, and ON-OFF gan-
glion cells. The neurons on their chip displayed properties that are central to biologi-
cal vision: receptive fields, spike coding, adaptation, band-pass filtering, and comple-
mentary signaling. Due to the lack of feedback from horizontal cells to photoreceptor
cells, as well as amacrine cells, their design could only demonstrate limited functions.
Hasegawa and Yagi emulated the architecture and functionality of the vertebrate outer
retina[27]. Their silicon retina carries out the spatial filtering of input images instan-
taneously, using embedded resistive networks that emulate the receptive field structure
of the outer retinal neurons, and a digital computer carries out temporal filtering of the
spatially filtered images to emulate dynamic properties of the outer retinal circuits. The
aim of their study was to emulate dynamic neural images produced by bipolar cells
in response to natural scenes in real time. Kameda and Yagi modeled the outer retina
using an analog neuromorphic multi-chip system to mimic the hierarchical structure of
the outer retina[37]. The functional network circuits were divided into two chips: the
photoreceptor network chip (P chip) and the horizontal cell network chip (H chip). The
output images of the P chip are transferred to the H chip using analog voltages through
the bus. An off-chip differential amplifier that models the bipolar cell layer takes input
from photoreceptors and horizontal cells. Their design realized a receptive field that
carries out smoothing and contrast enhancement on input images. Zaghloul and Boa-
hen proposed a silicon retina that reproduces the signals in the optic nerve[77]. Their
18
approach is to design an artificial retina from an information theory point of view. They
included both ON and OFF cone pathways in their model. Their model has the corre-
sponding nodes representing the responses of the retinal cells. The retinal cells in their
model include photoreceptors, horizontal cells, ON/OFF bipolar cells, wide/narrow field
amacrine cells, and 4 types of ganglion cells (namely ON-transient, ON-sustained, OFF-
transient, and OFF-sustained ganglion cells). The combination of all these retinal cells
in their transistor circuit reproduced the responses of ganglion cells. They tested the
silicon retina by applying impulses with different frequencies.
2.2.2 Neuromorphic Models of Motion Detection in the Visual Sys-
tem
Andreou and Strohbehn designed an analog VLSI processor for computer vision based
on the Hassenstein-Reichardt-Poggio model for information processing in the visual
system of a fly[2]. In their model, the delayed photoreceptor responses may correlate
with the current responses of neighboring receptors and perform directional selectivity.
Liu was also inspired by motion computation in the fly’s visual system and created a
neuromorphic circuit model of global motion processing in the fly[44]. Benson and
Delbruck proposed a silicon retinal model for direction selectivity[5]. Their design
used inhibitory connections in the null direction to perform the direction selectivity.
They included only photoreceptor cells and direction-selective ganglion cells (DSGCs)
in their model. Etienne-Cummings assumed primate motion detection is performed
in the cortex and analyzed insect and primate visual motion detection in a hardware
implementation[20]. However, the comparison does not include motion detection in
the retina. Wang and Liu designed an analog VLSI network using spiking neurons for
motion detection [69] which was based on the model proposed by Rao for explaining
19
the formation of direction- and velocity-selective cells in the visual cortex[64]. Never-
theless, their model was not focusing on the retina.
2.2.3 Implantable Artificial Retinas
Researchers led by Humayun on the ”Artificial Retina Project” are developing an
implantable microelectronic retinal prosthesis. Their approach is to develop an
implantable microeletronic retinal prosthesis that restores sight to people blinded by reti-
nal diseases. Visually-impaired patients whose conditions are not congenital (namely,
the optic nerve and visual cortex remain functional) can undergo a surgical procedure
involving surgically implanting a special microchip behind the retina to restore partial
sight [35]. Their most recent retinal prosthesis system is Argus II. It consists of five
main parts: (1) digital camera built into a pair of glasses which captures images in real
time and sends images to a microchip, (2) a video-processing microchip built into a
handheld unit that processes images into electrical pulses representing patterns of light
and dark and sends the pulses to a radio transmitter in the glasses, (3) a radio transmit-
ter that wirelessly transmits pulses to a receiver implanted above the ear or under the
eye, (4) a radio receiver that sends pulses to the retinal implant by a hair-thin implanted
wire, and (5) a retinal implant with an array of 60 electrodes on a chip measuring 1
mm by 1 mm. The camera is implanted in the frame of one’s eyeglasses to stimulate an
array of electrodes placed on the retinal surface. The moving images are sent along the
optic nerve to the brain. The device can provide sight (the detection of light) to people
who have become blind from degenerative eye diseases like macular degeneration and
retinitis pigmentosa. The ultimate goal of the project is to restore reading ability, facial
recognition, and unaided mobility for the blind [26]. Due to the vast knowledge required
from multiple sources, the ”Artificial Retina Project” requires multidisciplinary collab-
oration with medical science, material science, neuroscience, biomedical engineering,
20
and electrical engineering. Parker and Azar have proposed a bio-inspired 3D hierarchi-
cal pyramidal architecture for a synthetic retina[59]. They proposed that future artificial
retinas should maintain a hierarchical structure similar to that of the biological retina.
To substitute an artificial retina for a biological retina, an artificial retina should be able
to process visual information in a similar way as a biological retina would. In other
words, the future artificial retina should be equipped with the capability of processing
visual information including adaptation, dynamic behaviors, and extracting useful and
variant information from the scene. Their model aims to mimic the overall structure,
connectivity, and functionality of the human retina. They have also raised several chal-
lenges for future retinal prostheses which include power consumption, modeling visual
information processing, encoding visual information into spikes, and biocompatiblility.
2.3 Comparison of the Research to State of the Art in
Neuromorphic Designs of the Retina
For this dissertation research, we implement a portion of the starburst amacrine cell
(SAC) and differential motion detection model found the vertebrate retina. Modeling
this aspect of the vision system has not been done by others and may be useful for ser-
vice robots, autonomous vehicles and other applications that require processing dynamic
visual information in real time.
Moreover, we use the circuits to validate the importance of the feedback and lateral
pathways in performing motion sensing functions of a silicon retina. In order to achieve
this, we have implemented many CMOS circuits that model the mechanisms in the bio-
logical retinas. The mechanisms that we implemented can be removed by disconnecting
feedback or lateral connection among CMOS retinal cells. This allows us to perform
21
demonstrations by comparing the silicon retina tested to one altered by removing these
pathways and observing how the behaviors in the silicon retina are changed.
Table 2.1 compares our research to the related work that we have described.
22
Name Target Approach Input Size Output Functionality
BioRC Retina
(2012)
Inner and Outer
Retina (the
Vertebrate
Retina)
CMOS transistors
490
photoreceptors
Spikes
(1)Contrast Enhancement
(2)Center-Surround Antagonism
(3)Directional Selectivity
(4)Differential Motion Detection
Mahowald(1992) Outer Retina CMOS transistors
4096
photoreceptors
Graded
Potential
(1)Contrast Enhancement
(2)Center-Surround Antagonism
Hasegawa and
Yagi(2008)
Outer Retina Digital Processing
(PC) and FPGA
1840 photore-
ceptors
Graded
Potential
Spatiotemporal Responses of
Bipolar Cells
Kameda and
Yagi(2006)
Outer Retina FPGA
1840
photoreceptors
Graded
Potential
(1)Contrast Enhancement
(2)Center-Surround Antagonism
Zaghloul and
Boahen (2001)
Inner and Outer
Retina
CMOS transistors 5760 photo-
transistors
Spikes Responses over Different Fre-
quencies
Delbruck and Liu
(2004)
Inner and Outer
Retina w/o
Amacrine Cells
CMOS transistors
7
photoreceptors
Spikes
(1)Contrast Enhancement
(2)Center-Surround Antagonism
Andreou and
Strohbehn (1990)
Fly’s Visual Sys-
tem
CMOS transistors 50 photore-
ceptors
Graded
Potential
Directional Selectivity
Benson and Del-
bruck (1997)
Photoreceptors
and Direction
Selectivity (DS)
Cell
CMOS transistors 1927 photore-
ceptors
Graded
Potential
Directional Selectivity
USC Artificial
Retina Project
(2012)
Implantable Arti-
ficial Retina
Microelectronic
Prosthesis
60 electrodes Current Restore Sight to the Blind
Table 2.1: Comparison to State of the Art Research in Neuromorphic Designs of the Retina
23
Chapter 3
The Outer Retina Design
This chapter presents the outer retinal circuits as well as their testing results. The outer
retina has been modeled by others extensively. In order to be compatible with the inner
retinal circuits in Chapter 4, we built our own outer retinal circuits that perform edge
detection, contrast enhancement, and center-surround antagonism. We also create an
external control knob (i.e. glutamate reuptake) to control the amount of neurotransmitter
from the photoreceptor reaching a post-synaptic site of the horizontal cell. If there are
any additional mechanisms that control the glutamate reuptake found in the future, we
may include them in our model and emulate the corresponding impacts to the visual
processing in the outer retina. The circuit models include the photoreceptor circuit, the
horizontal cell compartment circuits, and the transient on-type bipolar cell circuit.
3.1 The Photoreceptor Circuit and Testing Results
The photoreceptor design is a modified version of the adaptive photoreceptor design
by Delbruck and Mead[13] shown in Figure 3.1. With a delayed feedback path, their
adaptive photoreceptor provides high gain for transient signals that are centered around
the adaptation point. Due to its simple structure and the adaptive property, we used their
design and made some changes to meet our requirements, namely modeling the output
response of biological photoreceptors. The modifications are as follows: (1) the design
was revised to better match the responses of biological photoreceptors, (2) a transistor,
whose gate receives feedback from the horizontal cell (i.e. HC fb) was added, (3) the
24
Figure 3.1: The Photoreceptor Circuit Design By Delbruck and Mead
transistors M1 and M2 were added for adjusting the light sensitivity. Please refer to
Figure 5.1. When absorbing photons, the photodiode induces a current flow that pulls
down the gate voltage of M3. It results in an increase of the gate voltage of M9. And,
the output, therefore, drops. After some delay that is produced by M8 and C1, the output
response will lower the gate voltage of M6 that pulls down the gate voltage of M9. As a
result, the output increases. This adaptive property is due to the negative feedback loop
that models the adaptive property of the biological photoreceptor. We can control the
sensitivity of the photoreceptor circuit by adjusting the sensitivity control knob. If the
voltage level of sensitivity is high, transistor M1 is more resistive and the gate voltage of
M3 can be pulled up higher, given the amount of current flowing through the photodiode.
If the voltage level of sensitivity is low, transistor M1 is less resistive and requires a
stronger light, namely more current flowing through the photodiode, to pull up the gate
voltage of M3 in order to evoke a stronger response at the output.
25
Figure 3.2: The Photoreceptor Circuit Design
The simulations of the photoreceptor circuit in Figure 3.3 and Figure 3.4 were con-
ducted using TSMC 18 CMOS (180nm) technology in SPECTRE and sensitivity was
set to be 0.9 V olts. (We will explain how changing sensitivity alters the output response
later.) We used a current source to model the photocurrent flowing through the photo-
diode. In Figure 3.3, we observed the output response during the time window from 0
to 400 us. We firstly applied 200nA as initial DC input current and changed the current
intensity at time 100 us. We tried different amounts of input changes (200nA, 220nA,
240nA, 260nA, 280nA, 300nA, and 320nA). As the amount of input change increases,
the response of the photoreceptor output increases as well. The overshoots after removal
of light are due to the internal feedback of the photoreceptor design. Similar responses
are observed in biological photoreceptors as well [9].
26
Figure 3.3: Simulation Result of an Isolated Photoreceptor
Figure 3.4 shows the relationship between the input range and the output range
under 13 different amounts of initial photocurrent (including 100nA, 500nA, 1000nA,
1500nA, 2000nA, 2500nA, 3000nA, 3500nA, 4000nA, 6000nA, 8000nA, 10000nA,
and 15000nA). The blue trace (labeled by DC) crossing all the other curves repre-
sents the output voltage level of the photoreceptor at different amounts of initial pho-
tocurrent before changing the photocurrent. Under different amounts of initial pho-
tocurrent, we varied the photocurrent intensity, recorded the peak values of the output
responses, and formed 13 different curves labeled by their values of initial photocur-
rent (100nA, 500nA, 1000nA, 1500nA, 2000nA, 2500nA, 3000nA, 3500nA, 4000nA,
27
Figure 3.4: Response Ranges of the Photoreceptor (When Sensitivity is 900 mV)
6000nA, 8000nA, 10000nA, and 15000nA) shown on the right side of the figure. As
the amount of initial photocurrent increases, the input range to cause responses also
increases. In other words, the photoreceptor is sensitive to a small input change at a
relatively dark background while less sensitive to a small input change at a relatively
bright background.
We examined the effect of changing the sensitivity of the photoreceptor. We first
gave 200 nA of photocurrent as the initial value. Then, we changed the photocurrent
and observed the amplitude of the output response under seven different values of the
sensitivity input (including 0.6V , 0.5V , 0.4V , 0.36V , 0.34V , 0.32V , and 0.3V). We used
the photocurrent intensity as the x-axis and the response amplitude as the y-axis to plot
Figure 3.5. In this figure, eight different curves were labeled by their values of the
sensitivity input. We observed that the input range (i.e. the range of the photocurrent
intensity) is shifting when changing the sensitivity input. Therefore, we concluded that
28
Figure 3.5: The Range of the Responses of an Isolated Photoreceptor Given Different
Sensitivity Values Ranging from 0.6 to 0.3 V olt
the photoreceptor requires a stronger input changes to evoke the output response as the
sensitivity input decreases,.
To better analyze the circuit’s behavior, we further found the best-fit line (in Figure
3.6) illustrating the output response and compared it with the response of the photore-
ceptor circuit. The equation of the best-fit line is shown in Equation 3.1 in which we
decomposed the equation into three components, i.e. DC, pull-down, and pull-up.
29
Figure 3.6: Best-Fit Line Representing the Response of the Photoreceptor Circuit
Response(t)=0:0689(0:4462e
0:01937t
)+(0:48e
0:1062t
)(Volts) (3.1)
DC (i.e. the first term in Equation 3.1) helps to adjust the DC level of the output.
The pull-down component (i.e. the second term in Equation 3.1) corresponds to the pho-
totransduction process in the biological photoreceptor that decreases the output voltage
when light is injected. The pull-up component (i.e. the last term in Equation 3.1) cor-
responds to the light adaptation process in the biological photoreceptor that pulls the
output back to its original DC level. At time 0, these three components determines the
resting potential of the output. To better visualize each component, we plotted them in
Figure 3.7. Note that we divided the DC component into two parts and merged them
into the pull-up and pull-down components, allowing us to easily observe the curves.
30
Figure 3.7: Curves Representing the Decomposed Terms of the Best-Fit Line
Moreover, we collected the output responses of biological photoreceptors recorded
from turtles[9], humans[57], and tiger salamanders[23]. We compared them with the
circuit’s outputs in Figure 3.8. We concluded that the photoreceptor circuit is about
4000-10000 times faster than biological photoreceptors and the amplitude of the circuit
output is about 20-50 times larger than that of biological photoreceptors.
31
Figure 3.8: Comparison of the photoreceptor Output to Biological Photoreceptors
32
3.2 Horizontal Cell Design
The horizontal cell compartment circuit consists of a two-stage non-inverting amplifier
that mirrors the response of the input and an internal delay element, a diode-connected
PMOS transistor, between two amplifiers, as shown in Figure 3.9. To model the horizon-
tal cell layer, the output, HC membrane potential, is connected to other compartments
through pass transistors that are used to model resistors. The pass transistors act like
resistors, which has been determined to be a good model for horizontal cells [67][75].
Figure 3.9: The Horizontal Cell Compartment Design
33
3.3 The Outer Retina Network and Testing Results
In Figure 3.10, we show the circuit implementing the interaction of one photoreceptor
and one horizontal cell (HC) compartment. Between two circuits, we inserted a buffer
with a control knob, glutamate reuptake, that models the effect of glutamate reuptake.
The input termed ”glutamate reuptake” can be used to demonstrate the impact of glu-
tamate uptake on the horizontal cell. The outputs of the photoreceptor and glutamate
reuptake control determine the voltage response of glutamate release reaching the post-
synaptic site of the horizontal cell. The response of the HC compartment mirrors the
response of ”glutamate concentration” with the internal delay modeled by transistor
M13, and is modulated by the output responses of its neighboring HC compartments.
When the voltage of photoreceptor’s output drops, the HC compartment’s output drops
as well. The response of each HC compartment feeds back to the photoreceptor and
helps to perform some functions such as contrast enhancement and center-surround
property. As we explained in Chapter 2, the feedback from the HC compartment to
the photoreceptor is negative. To demonstrate the functions of the feedback, we will
compare the outer retina design tested to one altered by removing the horizontal cell
layer in this section.
34
Figure 3.10: The Interaction Between a Photoreceptor Design and a Horizontal Cell
Compartment Design
To explain the connections of the outer retina, a one-dimensional outer-retinal net-
work structure is shown in Figure 3.11. One photoreceptor connects to one horizontal
cell compartment. A complete horizontal cell is formed by connecting a few compart-
ments with pass transistors. Each photoreceptor influences others through the horizon-
tal cell compartments and the resistor network. To demonstrate the properties of the
outer retina network, we constructed a two-dimensional network structure and did some
experiments which will be shown in the following section.
35
Figure 3.11: A One-Dimensional Photoreceptor-Horizontal Cell Network
36
3.3.1 Glutamate Reuptake
As we described in the previous section, we inserted a buffer with a control knob, glu-
tamate reuptake, that models the effect of glutamate reuptake. The rate of glutamate
reuptake is represented by the voltage level of the input, glutamate reuptake. To exam-
ine changing the input of glutamate reuptake, we used the circuit shown in Figure 3.10
that contains a photoreceptor circuit and a HC compartment circuit. We gave 200 nA
of photocurrent and varied the glutamate reuptake rate from low to high by sweeping
glutamate reuptake from -0.9 to 0.9 V olt in the circuit. We observed the change of gluta-
mate concentration, along with the output DC levels of the photoreceptor and horizontal
cell. Note that the voltage level of glutamate concentration corresponds to the glutamate
concentration in the synaptic cleft that. The higher the voltage level, the higher the glu-
tamate concentration. And, we used the glutamate reuptake rate as the x-axis and the
output DC level as the y-axis to plot the results shown in Figure 3.12. As the glutamate
reuptake rate increases, the DC level of glutamate concentration will reduce. The DC
level of the horizontal cell output will decrease due to less excitation provided by glu-
tamate. The DC level of the photoreceptor will increase due to less inhibition from the
horizontal cell.
37
Figure 3.12: Simulation Result for Changing Glutamate Reuptake from -0.9 to 0.9 V olt
3.3.2 Contrast Enhancement
To demonstrate that the interaction of photoreceptors and horizontal cells can enhance
the contrast of two inputs, we use two configurations to perform the experiments as
shown in Figure 3.13: One has two photoreceptors; the other has two photoreceptors
interacting with a horizontal cell. We used eight different amounts of photocurrents
(i.e. 0, 100, 200, 300, 400, 500, 600, and 700 nA) as the background light intensities
to test the circuits. In each case, we increase the amount of photocurrent, measure the
difference of the two outputs, and calculate the amount of contrast enhancement by
using the following equation.
38
Contrast Enhancement =
The output dierence with feedback
The output dierence without feedback
(3.2)
Note that the output differences in both cases are measured from resting level to
peak. The simulation results of the eight cases are plotted in Figure 3.14, where eight
different amounts of photocurrents (i.e. 0, 100, 200, 300, 400, 500, 600, and 700 nA)
as the background light intensities are used respectively. In each plot, x- and y- axis
represent the two inputs and z-axis represents the amount of contrast enhancement. For
all the cases shown in the figure, the amount of contrast enhancement increases as the
amount of the two inputs are getting closer (namely along the diagonal line). However,
when the difference of the two inputs is really small, the amount of contrast enhance-
ment will quickly drop to one. Therefore, we can infer that the difference of the two
inputs must pass a certain amount of threshold to exhibit contrast enhancement. Based
on the simulation results in Figure 3.14, the threshold increases as the background light
intensity increases.
39
Figure 3.13: Two Configurations for Testing Contrast Enhancement
40
Figure 3.14: Simulation Results of Contrast Enhancement
41
To explain the contrast enhancement in the circuit, please refer to Figure 3.10. When
the light shines on the photodiode, the gate voltage of transistor M3 increases and
then the output of the photoreceptor decreases. The output of the horizontal cell also
decreases and averages with the responses of its neighboring horizontal cell compart-
ments. The averaging output response from the horizontal cell compartment then feeds
back to the gate of transistor M5 which pulls down the source voltage of transistor M3.
As you notice, transistor M3 whose conductivity is proportional to the light intensity is
competing against transistor M5 which receives the feedback from the horizontal cell
compartment. If the light intensity is weak, the feedback can easily increase the output
response of the photoreceptor and make the output more positive. If the light intensity is
strong, the feedback will not be able to increase the output voltage of the photoreceptor
too much. Due to this property, the circuit can exhibit contrast enhancement.
3.3.3 Center-Surround Property
In the previous section, we demonstrated that the neuromophic design can exhibit con-
trast enhancement. In this section, we constructed a larger network to further demon-
strate the center-surround property. As we explained in Chapter 2, horizontal cell (HC)
to cone feedback helps establish the center-surround arrangement of visual receptive
fields. We will demonstrate the complete retinal pathway performing center-surround
property. In this section, we observed the response of center photoreceptors instead. In
the case of dark center bright surround, the center photoreceptors receive a strong feed-
back from the surround photoreceptor due to the horizontal cells. Therefore, when the
light shines on the surrounding photoreceptors, one should expect that the center pho-
toreceptors will be depolarized due to the strong feedback from the horizontal cell. In
the case of a bright center and dark surrounding, the surrounding photoreceptors receive
42
a strong feedback from the center photoreceptors due to the feedback from the hori-
zontal cells. Therefore, when the light shines on the center photoreceptors, one should
expect that the center photoreceptors would be hyperpolarized due to the weak feedback
from the horizontal cells.
In order to demonstrate the center-surrounding property in the outer retinal circuit,
we constructed a 14 by 14 Photoreceptor-Horizontal Cell (PHC) array as shown in Fig-
ure 3.15. Each photoreceptor receives feedback from the horizontal cell compartment it
connects to. The output of each horizontal cell compartment is connected with its neigh-
boring compartments using pass transistors. To test the 14 by 14 PHC array, we use two
different input patterns, dark-center-bright-surround and bright-center-dark-surround.
The simulation results are shown in Figure 3.16. The left column and right column
show the case of dark-center-bright-surround and bright-center-dark-surround respec-
tively. We also measured the difference of the response across the edge. In each plot, X-
and Y-axis indicate the index along the x and y direction respectively. The location of
each photoreceptor can be decided by these two indexes. Z-axis represents the voltage
level of the corresponding photoreceptor. The six plots in Figure 3.16 are recorded at
the moment when the responses of the photoreceptors reach their maximum/minimum
value. In both cases (left column and right column), the differences in the case of having
the horizontal cell are 0.191 V olt and 0.192 V olt respectively while the difference in the
case of having no horizontal cell is only 0.15 V olt.
43
Figure 3.15: The Configuration of the 14 by 14 Photoreceptor-Horizontal Cell (PHC)
Array
44
Figure 3.16: Demonstration of the Center Surround Property
45
3.3.4 Edge Detection
We tested the 14-by-14 PHC array. Initially, the input photocurrents are all 200nA. At
time 150n sec, we injected an input stimulus which formed an edge as shown in Fig-
ure 3.17(a) and we observed the responses of the photoreceptors over time which were
plotted in Figure 3.17(b). The photoreceptors along the bright edge are strongly hyper-
polarized because they receive less feedback from the photoreceptors along the dark
edge. On the other hand, the photoreceptors along the dark edge receive strong feed-
back from the photoreceptors along the bright edge. It results in a strong depolarization
of the photoreceptors along the dark side.
46
Figure 3.17: The Responses of Photoreceptors to an Edge Over Time.
47
Chapter 4
The Inner Retina Design
This chapter contains the major results of the dissertation research, directional selectiv-
ity, differential motion detection, and the complete retinal pathway from photoreceptor
to ganglion-cell spiking. These mechanisms are demonstrated with neural compart-
ments that illustrate these particular behaviors, but entire cell designs for the amacrine
cells are not included in the research. Since the amacrine cells do not spike, the dendrites
designed could be joined easily into complete cells.
4.1 On-type Bipolar Cell Design
Bipolar cells relay the signals from the outer retina to the inner retina; they are consid-
ered an important stage from which segregation of visual signal occurs and initiate a
chain of parallel processing for higher visual areas in the brain [73]. Bipolar cells have
two types: ON bipolar cell and OFF bipolar cell. Both receive glutamate released from
photoreceptors but respond differently. ON bipolar cells are mediated by metabotropic
glutamate receptors (mGluR6) on the dendrite while OFF bipolar cells are mediated
by ionotropic glutamtate receptors (iGluRs). When light hits a photoreceptor, the pho-
toreceptor hyperpolarizes and causes less glutamate release. On bipolar cells react to
this change by depolarizing while OFF bipolar cells react by hyperpolarizing. ON type
bipolar cells have two sub-types: transient and sustained. Different types of bipolar cells
connect to specific types of ganglion cells to carry out various retinal computations[3].
48
Transient-ON bipolar cells are involved in the computation of differential motion
detection [4]. ON bipolar cells receive glutamate released from the photorecep-
tors. Recent studies have revealed that a Transient Receptor Potential-Like Channel
(TRP/TRPL Channel) is necessary for the depolarizing light response of ON-bipolar
cells, and further that the TRP channel is a component of the channel that generates this
light response [63] [49] [61]. The identification of TRPM1 in particular as the channel
gated by the mGluR6 signaling cascade in ON bipolar cells reveals a major role for TRP
channels in vertebrate vision[61]. Despite intensive research, the molecular nature of
the mGluR6-gated channel has remained elusive[61].
The mGluR6 cascade starts with the opening of the mGluR6 receptor through an
indirect metabotropic process and ends up with the closure of mGluR6-gated channel
(i.e. TRP channel) as shown in the left drawing of Figure 4.1. (The left part of Figure
4.1 is taken from [41].) This process enables the conversion of a sustained input from
photoreceptors into a more transient output. The response at the post-synaptic site is
called Light-evoked EPSP (L-EPSP). Like other neurons, the L-EPSPs on the dendrite
sum nonlinearly and produce a potential output at its axon terminal. Moreover, the
voltage-gated ion channels located on either their dendrites or somas can enhance the
bipolar cell L-EPSPs by amplifying the transient component of the response[33]. At
the axon terminal, one observes the output is rectified[16]. Each bipolar cell’s rectified
response ensures that its vote will be counted, because it cannot be vetoed by signals of
equal magnitude but opposite sign in other parts of the receptive field[16]. A simplified
transient-ON bipolar cell model is shown in Figure 4.3.
The circuit on the right shown in Figure 4.1, namely the post-synaptic circuit of
transient-ON bipolar cell, has negative feedback that controls the gate voltage of M2.
As a result, the L-EPSP output will be quickly pulled down as shown in the simulation
result in Figure 4.2. The ON BC synapse circuit models the mGluR6 cascade process
49
Figure 4.1: The Post-Synaptic Circuit of the Transient-ON Bipolar Cell
that converts a sustained input from photoreceptors into a more transient output, namely
L-EPSP. The output will be further processed to achieve various retinal computations in
the inner retina.
50
Figure 4.2: Simulation Result of the Post-Synaptic Circuit of the Transient-ON Bipolar
Cell
Figure 4.3: Simplified Transient-ON Bipolar Cell Model with Dendritic Computation
51
4.2 A Directionally-Selective Neuromorphic Cir-
cuit Based on Reciprocal Synapses in Starburst
Amacrine Cells
The starburst amacrine cell (SAC), found in the mammalian retina and with a char-
acteristic radially symmetric morphology, is thought to provide directional inhibitory
input to direction-selective ganglion cells (DSGCs)[76][1][22]. It is generally believed
that SACs first perform the neural computations that induce directional selectivity in
the ganglion cell. The computation of direction selectivity (DS) occurs at individ-
ual dendritic branches of each SAC and each dendritic branch acts as an independent
computation module[21]. Both the dendritic calcium signal and membrane voltage in
the dendritic tip generate a stronger response by the stimuli moving from the soma
towards the dendritic tip (namely centrifugal motion) than moving the opposite direc-
tion (namely centripetal motion)[21]. To explain the DS observed in the SACs, neuro-
scientists have proposed at least two fundamentally different mechanisms[28]: dendrite-
intrinsic electrotonics[60][66] and lateral inhibition[7][43]. Euler et al. demonstrated
that the intrinsic electrical mechanisms of SACs may produce DS without inhibitory
network interactions[28]. However, the lateral inhibition between two SACs enhances
the difference in response and generate a robust directional selectivity[43]. Moreover,
voltage gated channels (possibly Ca
2+
) being found in the distal dendrite [28] imply
that a super-linear summation may occur in the distal dendrite of the SAC.
The mechanisms thought to underlie are explained here. SACs receive glutamate
released from bipolar cells (BCs). Furthermore, the dendritic tip releases and receives
the GABA neurotransmitter. Euler et al. observed a weak DS at the soma but a strong
DS in the dendritic tips[21]. They also found that SACs have directional responses even
if GABA inhibitory interactions between the SACs are blocked pharmacologically[21].
52
Their results suggest DS in the starburst cell arises intrinsically from its distinctive
morphology. A centrifugal (CF) motion generates an in-phase response which can be
summed effectively with the response in the distal compartment. However, centripetal
(CP) motion generates an out-of-phase response that cannot be summed effectively with
the response in the distal compartment. Hence, SAC produces a stronger response in the
distal tip with respect to centrifugal motion.
The lateral inhibition between two overlapping SACs makes the DS response more
robust[43]. The distal dendrite of SACs releases GABA neurotransmitters and GABA
receptors are also found in the SAC dendrites. Therefore, as long as the processes of
two neighboring starburst cells overlap, they are likely to form reciprocal connections.
The reciprocal synapse, which is a positive feedback loop, can enhance the difference
in responses between the two SACs. The interactions between SACs are illustrated
in Figure 4.4. When the light moves to BC2, the distal dendrite of SAC1 produces a
voltage response and releases more GABA which inhibits the response of distal dendrite
in SAC2. SAC2 in turn produces less GABA release which enhances the response of
the distal dendrite in SAC1.
53
Figure 4.4: The Inhibitory Interaction Between Two SACs
Figure 4.5 illustrates the morphology of the amacrine cell, taken from [21]. The
authors compartmentalize a SAC into 3 compartments in a branch: distal, intermediate,
and proximal compartments. Therefore, we modeled these three compartments in the
same fashion. The distal compartments receive glutamate release from bipolar cells as
well as GABA release from other SACs. To model the distal compartment, we need
to have two inputs: one for glutamate; the other for GABA. The basic distal compart-
ment design is presented in Figure 4.6. The gate of the lower NMOS, M2, connects
to GABA input from other SAC while the gate of the upper NMOS connects to gluta-
mate input from bipolar cell. The output represents the membrane potential of the distal
compartment of SAC. The circuit can model the responses caused by the changes of glu-
tamate and GABA. (Glutamate, an excitatory neurotransmitter, depolarizes the output
54
Figure 4.5: Starburst Amacrine Cell
while GABA, an inhibitory neurotransmitter, hyperpolarizes the output.) From transis-
tor M1’s perspective, the basic distal compartment is a source follower. From transistor
M2’s perspective, the basic distal compartment is a common source amplifier.
55
Figure 4.6: Basic Distal Compartment of the SAC
Figure 4.7 shows two overlapped branches of two simplified biological starburst
amacrine cells (SACs) and the corresponding circuit implementation. The top diagram
represents two SACs interacting through a reciprocal synapse. The bottom diagram
depicts the correspondence in our circuit implementation. The somatic compartments
and signal propagation toward the soma are not being modeled in our circuit. Each
branch of the SAC model consists of an intermediate compartment and a distal compart-
ment. (The somatic compartments is not included in our model.) Both compartments
receive glutamate inputs from the bipolar cells. The signals will first go through the
wave-shaping circuits that convert the glutamate input into cation concentration inside
the cell. The cation concentration at the intermediate compartment will propagate to the
distal compartment through a delay circuit and be summed with the cation concentration
at the distal compartment. A fully-functional SAC would have feedback path. We only
consider propagation towards the dendritic tip for this simplified SAC.
The summation is implemented by using a voltage adder. In Figure 4.7, the sub-
script of the parameters for the left SAC is 1 and the subscript of the parameters for
the right SAC is 2. Here, we only use the name of the parameters without the sub-
scripts to explain these parameters. The output of the voltage adder labeled [Cation]
56
represents intra-cellular cation concentration at the distal compartment. The membrane
potential of the distal compartment labeledV
d
is influenced by [Cation] and GABA IN.
GABA IN represents GABA from another SAC. GABA release is the voltage output
modulated by V
d
and GABA reuptake. The voltage adder circuit is a modified ver-
sion of Chaoui’s Circuit[10] and is capable of performing non-linear summations of
intra-cellular [Cation]. In the wave-shaping circuit, the rise of glutamate induces more
current to charge the output capacitor C quickly to V
glutamate
V
th
, where V
th
is the
threshold voltage of the pull up transistor in the wave-shaping circuit. The pull-down
transistor provides a resistive path for discharging the output capacitor C when gluta-
mate input decreases. Therefore, the wave-shaping circuit can produce an output with
smaller response than the input and longer duration. The delay circuit uses a current-
mirror structure to model the propagation delay along the branch of the SAC.
57
Figure 4.7: Two Overlapped Branches of Two Simplified Biological Starburst Amacrine Cells (SACs)
58
Next, we demonstrated a starburst amacrine cell model with a reciprocal synapse.
Figure 4.7 illustrates the scenario we set up to perform the experiments. Consider
the case in which [Cation]
2
remains the same and [Cation]
1
increases. At the out-
set, bothV
d1
and GABA release1 increase. The increase of GABA release1 pulls down
V
d2
and GABA release2. The decrease of GABA release2 pulls upV
d1
. As a result,V
d1
is
increased due to the positive feedback loop. During this operation, transistor M8 enters
the linear region as GABA IN2 increases. Therefore, the gain of the common source
amplifier consisting of M8 and M7 decreases. Meanwhile, transistor M2 soon enters
the subthreshold region that allows V
d1
to increase more quickly than transistor M1 is
in the saturation region. However, the amount of increase is limited by the decrease in
gain of the common source amplifier consisting of M8 and M7. Eventually, the whole
loop reaches a stable state without divergences or oscillations. V
d1
is still proportional
to the amplitude of[Cation]
1
. We may conclude that the SAC design with the recipro-
cal synapse still possesses the property of graded potential output and the operation is
stable.
The experiments were conducted using TSMC 18 CMOS (180nm) technology using
Cadence SPECTRE software simulating the configuration with two branches as shown
in Figure 4.7 and also the configuration with only one branch. To demonstrate the
dendrite-intrinsic electrotonics of the SAC design, we applied two kinds of moving stim-
uli, namely centripetal motion and centrifugal motion, to the configuration with only
one branch and measured the responses of the distal compartment for both cases. The
results are plotted in Figure 4.8. The black trace and purple trace represent the inputs
to the simulation that are the outputs from the bipolar cells connecting to the interme-
diate compartment and distal compartment respectively. The red trace is the response
of the distal compartment. The glutamate inputs are generated by the outer retina cir-
cuit that we made. The moving stimulus from the intermediate compartment to the
59
distal compartment will first evoke a response at the intermediate compartment. After
some delay, the signal reaches the distal compartment. Meanwhile, the moving stimulus
has reached the distal compartment and the evoked response can therefore be summed
with the response from the intermediate compartment. This results in a larger volt-
age response at the distal compartment than at the intermediate compartment. For the
opposite moving stimulus, the response cannot be optimally summed because the signal
from the intermediate compartment cannot reach the distal compartment on time. The
simulation results demonstrate that the stimulus moving centrifugally evokes a stronger
response than moving centripetally.
To demonstrate the lateral inhibition between the SACs, which enhances the dif-
ference in response of the distal dendrite, we tested the two configurations described
before and compared their responses. We applied a stimulus moving centrifugally to
both configurations. We measured the responses of the distal compartments for both
configurations that are plotted in Figure 4.9. The black trace and purple trace in the
upper figure represent the inputs to the simulation that are the outputs from the bipolar
cells connecting to the intermediate compartment and distal compartment respectively.
The blue trace and red trace represent the responses of the distal compartment of with a
reciprocal synapse and having no reciprocal synapse respectively. The results indicate
that the positive feedback of the reciprocal synapse effectively enhances the response of
the distal compartment.
60
Figure 4.8: Simulation Results of a Single SAC with Respect to Both Centripetal and
Centrifugal Motion
61
Figure 4.9: Comparison of the Simulation Results Both with and without Reciprocal
Synapse to Centrifugal Motion
62
To characterize the circuit’s behaviors, we applied a moving stimulus at different
speeds and different input intensities. We observed the responses of the distal compart-
ment and plotted the simulation results in Figure 4.10 and Figure 4.11 respectively. Both
figures record the maximum amplitude of the response. In Figure 4.10, the responses
are measured by using a stimulus with photocurrent of 250 nA and the speed (as shown
in the x-axis) has been normalize on a scale from 1 through 100 . For centripetal (CP)
motion, the evoked responses are not sensitive to the speed in either case of having a
reciprocal synapse or having no reciprocal synapse. For centrifugal (CF) motion, the
evoked response is small when the speed is slow and fast. The results imply that CF
motion evokes a larger response than CP motion within a range of speed and the ampli-
tude is enhanced by the presence of the reciprocal synapse.
Figure 4.10: Amplitude of the SAC Distal Compartment vs. Speed of the stimulus
63
In Figure 4.11, the simulations are conducted by sweeping the CF motion with dif-
ferent amounts of photocurrents as the input along the centrifugal direction. We used the
moving stimulus at two speeds, 1 and 10. In each case, we plotted two curves reflecting
the presence and no presence of a reciprocal synapse. Given the presence of a reciprocal
synapse and the moving stimulus at a proper speed (i.e. 10), the response of the distal
compartment is stronger than other cases. Moreover, the response is enhanced by the
presence of the reciprocal synapse across the entire input range that we swept.
Figure 4.11: Amplitude of the SAC Distal Compartment vs. Input Intensity
64
Next, we further quantified starburst DS by calculating a DS index (DSI) used in a
SAC model [66].
DSI =
V
cf
V
cp
V
cf
+V
cp
(4.1)
whereV
cf
andV
cp
are the peak responses measured at the distal compartment in the cen-
trifugal and centripetal directions, respectively. A DSI of 0 indicates no DS, 1 indicates
maximal DS with the preferred direction being centrifugal, and -1 indicates maximal
DS with a centripetal preference[66]. We plotted the results in Figure 4.12 in which DS
index is measured and calculated with respect to the moving stimulus at different speeds
(Given a photocurrent of 250nA as the input). The maximum DSIs occur for both curves
when the peak DSI speed is equal to the propagation speed of the moving stimulus (the
speed is about 10) and the presence of the reciprocal synapse leads to a higher DSI.
Figure 4.12: DS Index
65
4.3 A Neuromorphic Circuit that Computes Differential
Motion
Detecting moving objects in a moving background or a dynamic scene is essential to the
survival of some animals. Circuitry computing differential motion is found in the bio-
logical retina. An object-motion-sensitivity (OMS) ganglion cell remains silent under
global motion of the entire image but fires when the image patch in its receptive field
moves differently from the background. The differential motion neuromorphic circuit
that we built compares the motion speeds of the central receptive field and peripheral
receptive field. In this section, we demonstrate that there is a response if motion speeds
of the central and peripheral receptive fields are different. However, the response is
suppressed if motion speeds of the central and peripheral receptive fields are the same.
The OMS neurons are highly tuned to detect differential motion between the recep-
tive field center and the periphery. The polyaxonal amacrine cell appears to be a plau-
sible candidate to transmit inhibition from the background region [4]. The inhibition
signal may be derived from polyaxonal amacrine cells that inhibit the bipolar cell synap-
tic terminal, close to the site of transmission but at some electrotonic distance from the
soma. In the salamander, the bipolar cells involved in the computation of differential
motion detection are mainly transient OFF-type bipolar cells [4]. Transient ON-type
bipolar cells are believed to be involved in this computation as well [4][24]. The tran-
sient ON-type bipolar cell produces a positive response lasting a short duration of time
when increasing the light intensity. The transient OFF-type bipolar cell produces a pos-
itive response lasting a short duration of time when decreasing the light intensity. Com-
pared to the sustained bipolar cells, the transient bipolar cells have a shorter duration of
the response. The bipolar cells not only relay the visual information from photoreceptors
but also shape visual response before transmitting to the inner retina. The details of the
66
signaling cascade in the bipolar cells are given by several researchers [63], [49], [61].
The signaling cascade can be considered a feedback effect that enables the conversion
of a sustained input from photoreceptors into a more-transient output [40].
We constructed a neuromorphic circuit that computes retinal differential motion
(Figure 4.13). The network consists of a 7-by-70 photoreceptor array, a layer of hor-
izontal cells, 10 bipolar cells, and one sublinear voltage adder that models an amacrine
cell. The photoreceptors and the horizontal cell layer that perform contrast enhancement
have been presented in the previous chapter. In our circuit, each receptive field is cov-
ered by 5 bipolar cells. One bipolar cell connects postsynaptically with 5 photoreceptors
and the amacrine cell connects postsynaptically with 5 bipolar cells and presynaptically
to axonal terminals in 5 other bipolar cells belong to a different receptive field. The
amacrine cell is modeled by using a voltage adder circuit[10] that performs non-linear
summations as before. Modeling the ganglion cells is presented in the next section. In
this section, we used a sublinear voltage adder to sum up the bipolar cells’ responses
from the central receptive field and measured the output response of the summation of
the bipolar cells. Therefore, we may easily observe the effect of inhibition from the
peripheral receptive field under different cases. The bipolar cells that we modeled in the
network are transient ON-type bipolar cells in which the mGluR6 (glutamate receptor)
cascade causes conversion of a sustained input from photoreceptors into a more tran-
sient output [40]. The post-synaptic circuit of the transient-ON bipolar cell has been
presented in the previous section and shown in Figure 4.14. The L-EPSPs are summed
and rectified at the output. Transistor M6 pulls down the output voltage if the input from
an amacrine cell is asserted. Transistor M6 models the glycine receptor that receives
the inhibition from the amacrine cell. Transistor M7, rectifying the negative response,
ensures that each bipolar cell’s vote will be counted and cannot be vetoed by signals of
equal magnitude but opposite sign in other parts of the receptive field
67
Figure 4.13: A Neuromorphic Circuit that Performs Differential Motion
Figure 4.14: A Complete Bipolar Cell Circuit Model
68
The simulations were conducted with TSMC18 CMOS (180nm) technology using
the SPECTRE simulator. We used the circuit configuration shown in Figure 2.7 and
applied grating stimuli moving from the right to the left of the receptive field. Grating
stimuli consists of black and white bars that are represented by the photocurrent of
200nA and 250nA respectively. The black bars span four photoreceptors while the white
bars span only one photoreceptor. We tried two different cases of moving grating stimuli
(i.e. the same speed and different speeds). The results are shown in Figure 4.15. The
responses shown in Figure 4.15 record the summation of the bipolar cells’ responses
in the central receptive field during the first 8 msec. When the moving grating stimuli
are moving at the same frequency, the responses are much smaller than those moving
at different frequencies. The smaller response in the upper waveforms of Figure 4.15 is
due to the timely inhibition from the peripheral receptive field. The responses of both
cases are small at the beginning of the simulation (before 1m sec) because the moving
bars have not spanned the entire receptive field.
69
Figure 4.15: Simulation Results of Differential Motion Detection Circuit, Showing the
Summation of the Bipolar Cells’ Responses over Time as a grating is shifted
We simulated the circuit using different speed combinations of the central and
peripheral receptive field (both from 50KHz to 5KHz) and measured the maximum out-
put response during the first 10 msec. The black and white bars are represented by
the photocurrents of 200nA and 250nA respectively. The results are shown in Figure
4.16. The x-axis and y-axis represent the speeds of moving bars in the peripheral and
central receptive fields. The cells falling on the diagonal line are the cases having the
same speed of moving bars in the peripheral and central receptive fields. The responses
recorded from those cells are smaller due to the timely inhibition from the peripheral
receptive field that suppresses the responses of the bipolar cells in the central recep-
tive field. We also observed that the different speeds of moving bars may still produce
small responses in certain cases, i.e. the highlighted cells not falling on the diagonal
line. Those cases occur when the speed of the moving bars in the peripheral receptive
70
field is a multiple of that in the central receptive field or when the bars in the peripheral
receptive field are moving much faster than the bars in the central receptive field. In the
latter case, the cells close to the lower left corner of Figure 4.16, the inhibition from the
peripheral receptive field is too strong to generate the response at the output. The high-
lighted cells in Figure 4.16 indicate the responses less than 0.35 V . Hence, we concluded
that an output response above a certain value (i.e. 0.35 V in the case we presented in
Figure 4.16) indicates that the motions from the central receptive field and peripheral
receptive field are at different speeds. However, an output response below that certain
value does not necessarily imply the motions from the central receptive field and the
peripheral receptive field are at the same speed due to a few failing cases that we found.
Moreover, the amplitude of the output response does not indicate the magnitude of the
speed difference according to the results we observed.
71
Figure 4.16: Maximum Bipolar Cell Responses under Different Speed Combinations
72
We then fixed the input intensity of black bars and swept different input intensities
of white bars. We have tried two different combinations of the speeds, i.e. at the same
speed and at different speeds. For the case having the same speed, we used 20KHz for
both the central and peripheral receptive fields. For the case having different speeds, we
used 20KHz and 12.5KHz for the central and peripheral receptive field respectively. The
maximum output responses by the sweeping input photocurrent of the white bars from
200nA to 450nA are shown in Figure 4.17. The black bars are represented by injecting
a fixed photocurrent of 200nA. For the case having the same speed, we used 20K Hz for
both the central and peripheral receptive fields. For the case having different speeds, we
used 20K Hz and 12.5K Hz for the central and peripheral receptive field respectively.
Among the range we swept, the output responses are suppressed if the speeds are the
same.
Figure 4.17: Responses over Different Input Intensities to Moving Gratings
73
We repeated the simulations that we used for constructing Figure 4.16 by applying
different bar spacings and bar widths as inputs. To better visualize the output response
under different speed combinations, we plotted bar charts that record the responses of
each case shown in Figure 4.18, Figure 4.19, Figure 4.20, and Figure 4.21. In Figure
4.18, we varied the bar width from 1 to 4 given bar spacing of 5. In Figure 4.19, we
varied the bar spacing from 6 to 12 given bar width of 1. In Figure 4.20, we varied the
bar spacing from 6 to 12 given bar width of 2. In Figure 4.21, we varied the bar spacing
from 6 to 12 given bar width of 3. In these figures, we observed that (1) the diagonal
lines where the speeds are the same have lower response, (2) the average responses are
maximum when the bar spacing is 7 which is also the spacing of the bipolar cells, and (3)
the average response decreases when the bar spacing increases due to less input stimuli
across the receptive fields.
Figure 4.18: Responses in Bar Chart I: The Bar Spacing is 5 and Vary the Bar Width.
74
Figure 4.19: Responses in Bar Chart II: The Bar Width is 1 and Vary the Bar Spacing.
75
Figure 4.20: Responses in Bar Chart III: The Bar Width is 2 and Vary the Bar Spacing.
76
Figure 4.21: Responses in Bar Chart IV: The Bar Width is 3 and Vary the Bar Spacing.
77
Since we are interested in knowing how the outer retina circuits impact the compu-
tation, we removed the feedback from the horizontal cell layer to photoreceptors and
observed how the responses are altered. After disconnecting the feedback pathways
by removing the horizontal cell layer, we repeated the simulations that we used for
constructing Figure 4.18, Figure 4.19, Figure 4.20, and Figure 4.21 and compared the
simulation results. To more easily compare two sets of results, we converted these bar
charts into 2-dimensional plots and took the speed difference of each speed combination
as the X-axis and response as the Y-axis. The results are shown in Figure 4.22, Figure
4.23, Figure 4.24, and Figure 4.25 respectively. Note that HC denotes horizontal cell
layer. In Figure 4.22, we varied the bar width from 1 to 5 given the bar spacing of 5. In
Figure 4.23, we varied the bar spacing from 6 to 12 given the bar width of 1. In Figure
4.24, we varied the bar spacing from 6 to 12 given the bar width of 2. In Figure 4.25, we
varied the bar spacing from 6 to 12 given the bar width of 3. The blue traces indicate the
results with the presence of the feedback in the outer retina while the red traces indicate
the results without the presence of the feedback in the outer retina. From these figures,
we observed that the amplitude of the response does not vary dorectly with the different
speeds of two receptive fields. Also, we observed that the response with the horizontal
cell layer is higher than the responses without the horizontal cell layer in most cases.
78
Figure 4.22: Comparing Responses in Curves I: The Bar Spacing is 5 and Vary the Bar Width.
79
Figure 4.23: Comparing Responses in Curves II: The Bar Width is 1 and Vary the Bar Spacing.
80
Figure 4.24: Comparing Responses in Curves III: The Bar Width is 2 and Vary the Bar Spacing.
81
Figure 4.25: Comparing Responses in Curves IV: The Bar Width is 3 and Vary the Bar Spacing.
82
Next, we need to quantify the amount of performance enhancement. We made Equa-
tion 4.2 to quantify the performance enhancement.
Enhancement=
Diff
w=HC
Diff
w=oHC
Diff
w=oHC
(4.2)
,where Diff
w=HC
indicates the difference between the average response at the
same speed and at different speeds with the presence of the horizontal cell layer and
Diff
w=oHC
indicates the difference between the average response at the same speed
and at different speeds without the presence of the horizontal cell layer. If the value
of Diff
w=HC
is larger than that of Diff
w=oHC
, one may conclude that including the
horizontal cell layer helps the performance because the difference between the average
response at the same speed and at different speeds increases after including the horizon-
tal cell layer.
We then plotted the results in Figure 4.26, Figure 4.28, Figure 4.27, and Figure 4.29.
In Figure 4.26, we varied the bar width from 1 to 5 given the bar spacing of 5. In Figure
4.28, we varied the bar spacing from 6 to 12 given the bar width of 1. In Figure 4.27, we
varied the bar spacing from 6 to 12 given the bar width of 2. In Figure 4.29, we varied
the bar spacing from 6 to 12 given the bar width of 3. Note that HC denotes horizontal
cell layer.
For example, Figure 4.26 include four sets of results by injecting different bar spac-
ings and bar widths as inputs. In each set, the overlapping bar on the right is the case
with a horizontal cell layer while the one on the left is the case without a horizontal cell
layer. In each overlapping bar, the taller bar indicates the average response at the same
83
Figure 4.26: Enhancement I: The Bar Spacing is 1 and Vary the Bar Width.
speed while the shorter bar indicates the average response at different speeds. Each fig-
ure has a table in the bottom that calculates how much the performance is enhanced in
each case. The average amount of enhancement across the cases studied is about 27.1%.
84
Figure 4.27: Enhancement II: The Bar Width is 1 and Vary the Bar Spacing.
Figure 4.28: Enhancement III: The Bar Width is 2 and Vary the Bar Spacing.
85
Figure 4.29: Enhancement IV: The Bar Width is 3 and Vary the Bar Spacing.
86
Next, we analyze the hardware cost in terms of transistor counts. The total number of
transistors with horizontal cell layer is 9599 while the total number of transistors without
horizontal cell layer is 6246. Hence, we may conclude that to get the enhancement we
need to invest approximately a 33% increase in hardware cost.
4.4 Retinal Pathways
In this section, we modeled complete retinal pathways by modifying the BioRC spiking
circuits previously published [36][31]. In Figure 4.30 (a), the circuit produces one sin-
gle spike if the input voltage passes the threshold voltage of transistor M2. The circuit
formed by transistors M5-M12 produces a single spike at the positive edge of the input.
Transistor M1 and M2 are used to detect the input voltage. When the input passes the
threshold voltage of transistor M2, the drain voltage of M2 is pulled down to about 0 V .
The inverters, inv1, inv2, and inv3, are used to stabilize the input and form a stable pulse
for the rest of the circuit formed by transistors M5-M12. In Figure 4.30 (b), the circuit
produces tonic spikes if the input is 0V . If the input increases, the output generates spikes
at a faster rate. The circuit formed by transistors M13-M20 produces tonic spikes, i.e.
a oscillator. The original tonic spikeing circuit [36] is able to change spiking frequency
by adjusting the gate voltage of transistor M14. The spiking frequency is inversely pro-
portional to the gate voltage of transistor M14. Hence, we added an inverting amplifier
to make the spiking frequency proportional to the input level. If the input increases,
the spiking rate of the circuit increases, as compared to the spontaneous spiking rate.
Meanwhile, the output produces tonic spikes whose frequency is about 3.3 Khz when
there is no input, i.e. 0V .
87
Figure 4.30: Spiking Circuits
4.4.1 Object-Motion-Sensitivity (OMS) Ganglion Cell Pathway
An object-motion-sensitivity (OMS) ganglion cell remains silent under global motion of
the entire image but fires when the image patch in its receptive field moves differently
from the background [58], [4]. The OSM ganglion cell pathway performs differential
motion detection, which has been presented in the previous section. In this section, we
demonstrate the complete OSM ganglion cell pathway by adding the single-spiking cir-
cuit (Figure 4.30 (a)) to the output of the differential motion detection model. The com-
plete OSM ganglion cell circuit model consists of a voltage adder and a single-spiking
circuit. The voltage adder sums bipolar cell inputs. The single-spiking circuit produces
one single spike at a time if the input from the voltage adder passes the threshold voltage
of transistor M2.
The simulations were conducted with TSMC18 CMOS (180nm) technology using
Cadence SPECTRE simulator. We used the circuit configuration shown in Figure 4.13
plus the OSM ganglion cell circuit model and applied grating stimuli moving from the
right to the left to the receptive field. The grating stimuli consists of black and white bars
that are represented by a photocurrent of 200nA and 250nA respectively. The black bars
span four photoreceptors while the white bars span only one photoreceptor. We recorded
88
Figure 4.31: Measure the Outputs
the responses from the voltage adder’s output labeled by 1 and the OMS ganglion cell’s
output labeled by 2 shown in Figure 4.31. The simulation results are presented in Figure
4.32 and Figure 4.33. In Figure 4.32, we observed that the output remains silent when
both speeds are the same. In Figure 4.33, the output produces spikes when the speeds
are different. The circuit without the OMS ganglion cell circuit testing has been fully
described in the previous section. Hence, we only presented two speed combinations
that demonstrate the OMS ganglion cell pathway in this section.
89
Figure 4.32: Responses at the Same Speed
90
Figure 4.33: Responses at the Different Speeds
91
Comparing the behaviors of the biological OMS ganglion cell Pathway, we applied 5
periodic jittering bars to each receptive field in the circuit model. The results are shown
in Figure 4.34. Note that the neuronal recordings of the biological OSM ganglion cell
are taken from the paper [4].
Figure 4.34: Comparison to the Biological Results
92
4.4.2 On-Center Ganglion Cell Pathway
In the retina, bipolar cells and ganglion cells are organized in such a way that each
cell responses to a small circular patch of the retina that defines the receptive field.
The receptive field of the retinal ganglion cell consists of a roughly circular central
area and a surrounding ring. Retinal ganglion cells have two basic types of receptive
fields, on-center/off-surround and off-center/on-surround. The center and its surround
are always antagonistic and tend to cancel each other’s activity. When no light spikes on
the receptive field, a spontaneous (tonic) level of spiking activity is recorded from the
ganglion cell.
Figure 4.35 explains the operation of an on-center ganglion cell pathway. Note that
we only show three photoreceptors in the figure for simplicity. The actual number of
photoreceptors appearing in the receptive field of a biological retina may be much more
than what we show in the figure. In Figure 4.35 (a), the light shines on the center
photoreceptor. The center photoreceptor hyperpolarizes due to the light during t1-t2.
The on bipolar cell also hyperpolarizes in response to the center photoreceptor. The
on ganglion cell generates spikes during t1-t2 due to the hyperpolarization of the on
bipolar cell. In Figure 4.35 (b), the light shines on the surround photoreceptors. The
center photoreceptor depolarizes due to the inhibition from the surround photoreceptors
through the horizontal cell during t1-t2. The on bipolar cell also depolarizes in response
to the center photoreceptor. The on ganglion cell is inhibited from firing during t1-t2
due to the depolarization of the on bipolar cell.
93
Figure 4.35: On-Center Ganglion Cell Pathway
In this section, we constructed a on-center ganglion cell pathway from photorecep-
tors to the ganglion cell and demonstrated the responses in Figure 4.36 and Figure 4.37.
The photoreceptor array consists of 7-by-7 photoreceptors. The central 3-by-3 photore-
ceptors is the center receptive field and the rest of the photoreceptors are the surround
receptive field. The bipolar cell circuit model consists of a voltage adder and an invert-
ing amplifier shown in Figure 4.38 and connects with the central 3-by-3 photoreceptors.
The ganglion cell circuit is shown in Figure 4.30 (b). The simulations were conducted
with TSMC18 CMOS (180nm) technology using the SPECTRE simulator. In Figure
4.36, both receptive fields are dark before 0.1m second. At time 0.1m second, we shine
a light on the center receptive field till 0.5m second. After 0.5m second, both receptive
94
Figure 4.36: Simulation Result of On-Center Ganglion Cell Pathway I
fields remain dark till the end of the simulation. We observed the response of the bipo-
lar cell and ganglion cell over the time. The ganglion cell generates more spikes when
the light strikes the center receptive field. In Figure 4.37, both receptive fields are dark
before 0.1m second. At time 0.1m second, we shine a light on the surround receptive
field till 0.5m second. After 0.5m second, both receptive fields remain dark till the end
of the simulation. We again observed the response of the bipolar cell and ganglion cell
over the time. The ganglion cell is prevented from spiking during 0.1m-0.5m second
due to the inhibition from the surround receptive field and produces spikes vigorously
due to less inhibition (i.e. more excitation to the center receptive field) from surround
receptive field after 0.5m sec.
Note that the horizontal cell plays an important role in the on-center ganglion cell
pathway. Without the horizontal cell, the communication among photoreceptors cannot
happen. Hence, we cannot demonstrate the spiking behaviors without the horizontal
cell.
95
Figure 4.37: Simulation Result of On-Center Ganglion Cell Pathway II
Figure 4.38: On Center Bipolar Cell Circuit Model
96
4.4.3 Directionally-Selective Ganglion Cell Pathway
The Directionally-Selective Ganglion Cell (DSGC) is the output neuron that computes
motion direction in the retina. It codes motion direction by generating more spikes when
there is motion in a particular direction and generating less spikes (or no spikes) when
there is motion in the opposite direction. The underlying neural circuit and its operations
is presented in Figure 4.39. The starburst amacrine cell (SAC) responses differently with
respect to different moving stimuli and provides inhibition to DSGC while bipolar Cell
excites DSGC. In Figure 4.39 (a), A moving light from BC1 to BC2 produces a stronger
response in the dendritic tip of SAC which cancels the excitation from BC1. Hence,
DSGC does not produce spikes vigorously. In Figure 4.39 (b), the moving light from
BC2 to BC1 produces a weaker response in the dendritic tip of SAC which is not enough
to cancel the excitation from BC1. Hence, DSGC generates more spikes, as compared
to the tonic spiking rate.
Figure 4.39: Directionally-Selective Ganglion Cell Pathway
97
The simulations were conducted with TSMC18 CMOS (180nm) technology using
the SPECTRE simulator. The starburst amacrine cell (SAC) circuit model has been
explained in the previous section. The DSGC circuit model is presented in Figure 4.40.
DSGC circuit model consists of a voltage adder and the tonic-spiking circuit. The volt-
age adder takes input from the bipolar cell (BC) and starburst amacrine cell (SAC). Since
SAC inhibits DSGC, an inverting amplifier whose gain is about 1 is necessary.
Figure 4.40: Directionally-Selective Ganglion Cell (DSGC) Circuit model
98
The retinal network in Figure 4.41 and Figure 4.42 contains 2 bipolar cells (i.e. BC1
and BC2), one branch of SAC that has been presented in previous section, and a DSGC.
In Figure 4.41, the input stimulus moving from BC1 to BC2 evokes a stronger response
in the SAC that cancels out the excitation from BC1. Hence, the DSGC still produces
spikes tonically. In Figure 4.42, the input stimulus moving from BC1 to BC2 evokes a
weaker response in the SAC that is not enough to inhibit DSGC. Therefore, the DSGC
produces more spikes, as compared to the tonic spiking rate.
Figure 4.41: The Response of DSGC to Centrifugal Motion
99
Figure 4.42: The Response of DSGC to Centripetal Motion
100
Chapter 5
Conclusion and Future Research
5.1 Conclusion
The main goal of this research is to validate the importance of the feedback and lateral
pathways in a silicon retina and to predict the feasibility of CMOS Technology in imple-
menting a silicon retina. In this research, we demonstrated that some functions cannot
be achieved or performances degrade without feedback and lateral connections. Hence,
we concluded that it is worth incorporating feedback and lateral connections in the arti-
ficial retina even though it makes the retinal network complicated. As to the impact of
the feedback and lateral pathways, a network might result in exponential divergence or
exponential growth of oscillations if we incorporate feedback and lateral connections.
Hence, stability was an important consideration.
In Chapter 1, we addressed the motivation of pursuing this research. We further pro-
posed our hypothesis: Feedback and lateral interactions are important in a silicon retina
to model the response of a vertebrate retina. To address the hypothesis, we developed
circuit models that are presented in Chapter 3 and 4.
Chapter 2 started by providing a review of biological retinas that covers basic struc-
ture of biological retinas and some computational models found in biological retinas.
These computational models were implemented in circuits and presented in Chapter 3
and Chapter 4. We further compared the related retinal neuromorphic research with our
work in this chapter.
101
The retinal circuit models that show the importance of including feedback and lat-
eral pathways in the silicon retina are presented in Chapter 3 and Chapter 4. We have
designed and simulated several major types of neural circuits that model the retinal
neruons with TSMC 18 CMOS (180nm) technology using SPECTRE and particularly
modeled the responses of the intermediate retinal cells by considering their underly-
ing (both inter-cellular and intra-cellular) mechanisms. Our retinal design maintains a
hierarchical structure similar to that of the biological retina. Lateral connections with
horizontal cells and amacrine cells are employed, along the feedback within the inner
and outer plexiform layers of the retina. We validate the importance of the feedback
and lateral pathways in the silicon retina by comparing the silicon retina tested to one
altered by removing these pathways and observed how the behaviors in the silicon retina
are changed. By including these circuits, we are able to model the first-order function-
ality of a portion of the retina from photoreceptors to ganglion cells that maintains a
hierarchical structure similar to that of biological retina. The retinal pathways that we
modeled include an on-type ganglion cell pathway, directionally-selective ganglion cell
(DSGC) pathway, and object-motion-sensitivity (OMS) ganglion cell pathway. In con-
clusion, we demonstrate that feedback and lateral connections are essential for process-
ing visual information in our circuit model even though those connections complicate
the retinal circuit.
5.2 Future Research
No research can claim to perfectly provide solutions to all aspects of the problems
related to a topic. During the course of this research, numerous research problems were
encountered. Some of these problems were addressed in this thesis while others were
left as subjects of future research.
102
A complete silicon retina should perform many different functions. The course of
this research is too short to accomplish all the known functions of biological retinas
in CMOS circuit. In fact, there are even more unknown functions and cell types of
biological retina still requiring to be explored by biologists and neuroscientists. We
believe that these functions may be useful for service robots, autonomous vehicles and
other applications. Hence, we will work with neuroscientists closely and implement
other neuromorphic circuits that model the operations of other biological retinal cells.
Retinal glial cells are another important future research that we will definitely pur-
sue. Glial cells outnumber neurons in the central nervous system (CNS) by10 to 1.
Traditionally, glial cells were believed to provide only passive structural and metabolic
support for neurons. In recent years, neuroscientists studied Glial-neural communica-
tion by monitoring the effect of both intercellular and intracellular glial Ca
2+
waves
on the electrical activity of neighboring neurons [56] [54]. The studies revealed that
light-evoked activity in retinal neurons results in an increase in the frequency ofCa
2+
transients in M¨ uller cells [55] [53]. M¨ uller cells modulate the neural activity by releas-
ing gliotransmitters including glutamate, ATP, and D-serine, which are induced by intra-
cellular [Ca
2+
]. The ganglion cell hyperpolarization is mediated by ATP release from
the M¨ uller cells. Released ATP is rapidly converted to adenosine. Adenosine, in turn,
then activates A1 adenosine receptors on the ganglion cells, leading to the opening of
K+ channels. The opening of K+ channel permits the exit of potassium ions from the
ganglion cell. As a result, the ganglion cell hyperpolarizes and does not fire[65]. Neu-
roscientists also observed that glutamate uptake by M¨ uller cell may inhibit the response
of retinal ganglion cells[29]. A two-way communication between retinal neurons and
glial cells may exist and suggest that retinal glial cells contribute to information pro-
cessing in the retina[65]. As a result a conclusion was drawn that glial cells are capable
of modulating the electrical activity of neurons within the retina [56]. It is very likely
103
that the glial cells in the retina serve certain functions that help visual processing. We
will work with neuroscientists who study the glial cells in the retina and study whether
it is beneficial to implement glial cells in silicon retina.
Another major challenge that we did not address in this thesis is power consumption.
Our circuit models contain many analog circuits requiring DC current flowing all the
time. Those circuit models consume significant power. We know that a biological retina
contains hundreds of millions of retinal cells. Therefore, even if we can significantly
decrease the power consumption of each component, a complete silicon retina still may
consume huge amounts of power. Therefore, power consumption is another factor that
might limit the scale of the retinal network. Our future research plans to decrease the
power consumption of silicon retinal circuits and predict how power consumption might
limit the scale of the retinal network.
This thesis is mainly focusing on the silicon retina. Future researchers working on
artificial retinas may need to look toward other advanced technologies with smaller size,
less power consumption, and more connectivity.
104
References
[1] Franklin R. Amthor, Kent T. Keyser, and Nina A. Dmitrieva. Effects of the destruc-
tion of starburst-cholinergic amacrine cells by the toxin AF64A on rabbit retinal
directional selectivity. Visual neuroscience, 19(4):495–509, 2002.
[2] A. G. Andreou and K. Strohbehn. Analog VLSI implementation of the
Hassenstein-Reichardt-Poggio models for vision computation. In Systems, Man
and Cybernetics, 1990. Conference Proceedings., IEEE International Conference,
pages 707–710, 1990.
[3] Gautam B. Awatramani and Malcolm M. Slaughter. Origin of transient and sus-
tained responses in ganglion cells of the retina. J. Neurosci., 20(18):7087–7095,
September 2000.
[4] Stephen A. Baccus, Bence P. Olveczky, Mihai Manu, and Markus Meister. A
Retinal Circuit That Computes Object Motion. J. Neurosci., 28(27):6807–6817,
July 2008.
[5] Ronald G. Benson and Tobi Delbr¨ uck. Direction selective silicon retina that uses
null inhibition. In Advances in Neural Information Processing Systems 4, vol-
ume 4, pages 756–763, 1991.
[6] Kwabena A. Boahen and Andreas G. Andreou. A Contrast Sensitive Silicon Retina
with Reciprocal Synapses. In Advances in neural information processing systems,
volume 4, pages 764–772, 1991.
[7] Lyle J. Borg Graham and Norberto M. Grzywacz. A model of the directional selec-
tivity circuit in retina: transformations by neurons singly and in concert, pages
347–375. Academic Press Professional, Inc., San Diego, CA, USA, 1992.
[8] A. Bringmann, T. Pannicke, J. Grosche, M. Francke, P. Wiedemann, S. Skatchkov,
N. Osborne, and A. Reichenbach. M¨ uller cells in the healthy and diseased retina.
Progress in Retinal and Eye Research, 25(4):397–424, July 2006.
105
[9] D. A. Burkhardt. Light adaptation and photopigment bleaching in cone photore-
ceptors in situ in the retina of the turtle. J. Neurosci., 14(3):1091–1105, March
1994.
[10] H. Chaoui. CMOS analogue adder. Electronics Letters, 31(3):180–181, 1995.
[11] Paul B. Cook and John S. McReynolds. Lateral inhibition in the inner retina is
important for spatial tuningof ganglion cells. Nature Neuroscience, 1(8):714–719,
December 1998.
[12] T. Delbr¨ uck and C. A. Mead. An electronic photoreceptor sensitive to small
changes in intensity. Neural Information Processing Systems, pages 720–727,
1989.
[13] T. Delbruck and C. A. Mead. Adaptive photoreceptor with wide dynamic range.
In Circuits and Systems IEEE International Symposium, pages 339–342, 1994.
[14] T. Delbruck and D. Oberhoff. Self-biasing low power adaptive photoreceptor. In
Circuits and Systems IEEE International Symposium, pages IV–844–7, 2004.
[15] Tobi Delbr¨ uck and Shih-Chii C. Liu. A silicon early visual system as a model
animal. Vision research, 44(17):2083–2089, 2004.
[16] J. B. Demb, K. Zaghloul, L. Haarsma, and P. Sterling. Bipolar cells contribute to
nonlinear spatial summation in the brisk-transient (y) ganglion cell in mammalian
retina. The Journal of neuroscience : the official journal of the Society for Neuro-
science, 21(19):7447–7454, October 2001.
[17] M. A. Dyer and C. L. Cepko. Regulating proliferation during retinal development.
Nature reviews. Neuroscience, 2(5):333–342, May 2001.
[18] Erika D. Eggers and Peter D. Lukasiewicz. Gabaa, gabac and glycine receptor-
mediated inhibition differentially affects light-evoked signalling from mouse reti-
nal rod bipolar cells. The Journal of Physiology, 572(1):215–225, April 2006.
[19] Germ´ an A. Enciso, Michael Rempe, Andrey V . Dmitriev, Konstantin E. Gavrikov,
David Terman, and Stuart C. Mangel. A model of direction selectivity in the star-
burst amacrine cell network. Journal of computational neuroscience, 28(3):567–
578, June 2010.
[20] R. Etienne-Cummings. Biologically Inspired Visual Motion Detection in VLSI.
International Journal of Computer Vision, pages 175–198, September 2001.
[21] Thomas Euler, Peter B. Detwiler, and Winfried Denk. Directionally selective cal-
cium signals in dendrites of starburst amacrine cells. Nature, 418(6900):845–852,
August 2002.
106
[22] Shelley I. Fried, Thomas A. M¨ unch, and Frank S. Werblin. Mechanisms and cir-
cuitry underlying directional selectivity in the retina. Nature, 420(6914):411–414,
November 2002.
[23] L. Gaal, B. Roska, S. A. Picaud, S. M. Wu, R. Marc, and F. S. Werblin. Postsy-
naptic response kinetics are controlled by a glutamate transporter at cone photore-
ceptors. Journal of neurophysiology, 79(1):190–196, January 1998.
[24] Maria N. Geffen, Saskia E. J. de Vries, and Markus Meister. Retinal ganglion cells
can rapidly change polarity from off to on. PLoS Biol, 5(3):e65+, March 2007.
[25] Tim Gollisch and Markus Meister. Eye smarter than scientists believed: Neural
computations in circuits of the retina. Neuron, 65(2):150–164, January 2010.
[26] Scott H. Greenwald, Alan Horsager, Mark S. Humayun, Robert J. Greenberg,
Matthew J. McMahon, and Ione Fine. Brightness as a function of current amplitude
in human retinal electrical stimulation. Invest. Ophthalmol. Vis. Sci., 50(11):5017–
5025, November 2009.
[27] Jun Hasegawa and Tetsuya Yagi. Real-time emulation of neural images in the outer
retinal circuit. The Journal of Physiological Sciences, 58(7):507–514, 2008.
[28] Susanne E. Hausselt, Thomas Euler, Peter B. Detwiler, and Winfried Denk.
A dendrite-autonomous mechanism for direction selectivity in retinal starburst
amacrine cells. PLoS biology, 5(7), July 2007.
[29] Matthew H. Higgs and Peter D. Lukasiewicz. Glutamate uptake limits synaptic
excitation of retinal ganglion cells. J. Neurosci., 19(10):3691–3700, May 1999.
[30] Hajime Hirasawa and Akimichi Kaneko. pH Changes in the Invaginating Synaptic
Cleft Mediate Feedback from Horizontal Cells to Cone Photoreceptors by Mod-
ulating Ca2+ Channels. The Journal of General Physiology, 122(6):657–671,
December 2003.
[31] Chih-Chieh Hsu, Alice C. Parker, and Jonathan Joshi. Dendritic computations,
dendritic spiking and dendritic plasticity in nanoelectronic neurons. In 2010 53rd
IEEE International Midwest Symposium on Circuits and Systems, pages 89–92.
IEEE, August 2010.
[32] Hain-Ann Hsueh, Alyosha Molnar, and Frank S. Werblin. Amacrine-to-amacrine
cell inhibition in the rabbit retina. J Neurophysiol, 100(4):2077–2088, October
2008.
[33] Tomomi Ichinose, Colleen R. Shields, and Peter D. Lukasiewicz. Sodium chan-
nels in transient retinal bipolar cells enhance visual responses in ganglion cells.
107
The Journal of neuroscience : the official journal of the Society for Neuroscience,
25(7):1856–1865, February 2005.
[34] Skyler L. Jackman, Norbert Babai, James J. Chambers, Wallace B. Thoreson, and
Richard H. Kramer. A Positive Feedback Synapse from Retinal Horizontal Cells
to Cone Photoreceptors. PLoS Biol, 9(5):e1001057+, May 2011.
[35] Michael Javaheri, David S. Hahn, Rohit R. Lakhanpal, James D. Weiland, and
Mark S. Humayun. Retinal prostheses for the blind. Annals of the Academy of
Medicine, Singapore, 35(3):137–144, March 2006.
[36] Jonathan Joshi, Alice C. Parker, and Chih-Chieh C. Hsu. A carbon nanotube spik-
ing cortical neuron with tunable refractory period and spiking duration. IEEE Latin
American Symposium on Circuits and Systems (LASCAS), 2010.
[37] S. Kameda and T. Yagi. An analog silicon retina with multichip configuration.
IEEE Transactions on Neural Networks, 17(1):197–210, January 2006.
[38] Maarten Kamermans and Iris Fahrenfort. Ephaptic interactions within a chemical
synapse: hemichannel-mediated ephaptic inhibition in the retina. Current opinion
in neurobiology, 14(5):531–541, October 2004.
[39] Maarten Kamermans, Iris Fahrenfort, Konrad Schultz, Ulrike Janssen-Bienhold,
Trijntje Sjoerdsma, and Reto Weiler. Hemichannel-Mediated Inhibition in the
Outer Retina. Science, 292(5519):1178–1180, May 2001.
[40] Tejinder Kaur and Scott Nawy. Characterization of Trpm1 desensitization in ON
bipolar cells and its role in downstream signalling. The Journal of Physiology,
590(1):179–192, January 2012.
[41] Chieko Koike, Takehisa Obara, Yoshitsugu Uriu, Tomohiro Numata, Rikako
Sanuki, Kentarou Miyata, Toshiyuki Koyasu, Shinji Ueno, Kazuo Funabiki, Akiko
Tani, Hiroshi Ueda, Mineo Kondo, Yasuo Mori, Masao Tachibana, and Takahisa
Furukawa. TRPM1 is a component of the retinal ON bipolar cell transduction
channel in the mGluR6 cascade. Proceedings of the National Academy of Sci-
ences, 107(1):332–337, January 2010.
[42] Helga Kolb. Glial cells of the retina, April 2007.
[43] Seunghoon Lee and Z. Jimmy Zhou. The synaptic mechanism of direction selectiv-
ity in distal processes of starburst amacrine cells. Neuron, 51(6):787–799, Septem-
ber 2006.
[44] Shih-Chi Liu. A neuromorphic aVLSI model of global motion processing in the
fly. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 47(12):1458–1467, Dec 2000.
108
[45] D. M. MacKay. Elevation of Visual Threshold by Displacement of Retinal Image.
Nature, 225(5227):90–92, January 1970.
[46] Misha A. Mahowald. VLSI analogs of neuronal visual processing: a synthesis of
form and function. PhD thesis, California Institute of Technology, Pasadena, CA,
USA, 1992.
[47] C. Mead and M. Mahowald. A silicon model of early visual processing. Neural
Networks, 1(1):91–97, 1988.
[48] Alyosha Molnar and Frank Werblin. Inhibitory feedback shapes bipolar cell
responses in the rabbit retina. J Neurophysiol, 98(6):3423–3435, December 2007.
[49] Catherine W. Morgans, Jianmei Zhang, Brett G. Jeffrey, Steve M. Nelson, Neal S.
Burke, Robert M. Duvoisin, and R. Lane Brown. Trpm1 is required for the depo-
larizing light response in retinal on-bipolar cells. Proceedings of the National
Academy of Sciences, 106(45):19174–19178, November 2009.
[50] T. A. Muench and F. S. Werblin. Symmetric interactions within a homogenous
starburst cell network can lead to robust asymmetries in starburst dendrites. Invest.
Ophthalmol. Vis. Sci., 45(5):4274+, May 2004.
[51] Thomas A. Munch, Rava A. da Silveira, Sandra Siegert, Tim J. Viney, Gautam B.
Awatramani, and Botond Roska. Approach sensitivity in the retina processed by a
multifunctional neural circuit. Nature Neuroscience, 12(10):1308–1316, Septem-
ber 2009.
[52] M. Murakami, Y . Shimoda, K. Nakatani, E. Miyachi, and S. Watanabe. GABA-
mediated negative feedback from horizontal cells to cones in carp retina. The
Japanese journal of physiology, 32(6):911–926, 1982.
[53] E. A. Newman. Glial modulation of synaptic transmission in the retina. Glia,
47(3):268–274, 2004.
[54] Eric A. Newman. Propagation of intercellular calcium waves in retinal astrocytes
and m¨ uller cells. J. Neurosci., 21(7):2215–2223, April 2001.
[55] Eric A. Newman. Calcium increases in retinal glial cells evoked by light-induced
neuronal activity. The Journal of neuroscience : the official journal of the Society
for Neuroscience, 25(23):5502–5510, June 2005.
[56] Eric A. Newman and Kathleen R. Zahs. Modulation of neuronal activity by glial
cells in the retina. J. Neurosci., 18(11):4022–4028, June 1998.
109
[57] Mahito Ohkuma, Fusao Kawai, Masayuki Horiguchi, and Ei-Ichi Miyachi. Patch-
clamp Recording of Human Retinal Photoreceptors and Bipolar Cellsy. Photo-
chemistry and Photobiology, 83(2):317–322, 2007.
[58] Bence P. Olveczky, Stephen A. Baccus, and Markus Meister. Segregation of object
and background motion in the retina. Nature, 423(6938):401–408, May 2003.
[59] Alice C. Parker and Adi N. Azar. A hierarchical artificial retina architecture. In
´
Angel B. Rodr’ıguez V´ azquez, editor, Bioengineered and Bioinspired Systems IV,
volume 7365, pages 736503+. SPIE, 2009.
[60] R. R. Poznanski. Modelling the electrotonic structure of starburst amacrine cells
in the rabbit retina: a functional interpretation of dendritic morphology. Bulletin
of mathematical biology, 54(6):905–928, November 1992.
[61] Christophe Ribelayga. Vertebrate vision: Trp channels in the spotlight. Curr Biol,
20(6):R278–R280, March 2010.
[62] Botond Roska and Frank Werblin. Rapid global shifts in natural scenes block
spiking in specific ganglion cell types. Nature Neuroscience, 6(6):600–608, May
2003.
[63] Yin Shen, J. Alexander Heimel, Maarten Kamermans, Neal S. Peachey, Ronald G.
Gregg, and Scott Nawy. A transient receptor potential-like channel mediates
synaptic transmission in rod bipolar cells. The Journal of neuroscience : the offi-
cial journal of the Society for Neuroscience, 29(19):6088–6093, May 2009.
[64] A. P. Shon, R. P. Rao, and T. J. Sejnowski. Motion detection and prediction through
spike-timing dependent plasticity. Network (Bristol, England), 15(3):179–198,
August 2004.
[65] Larry R. Squire. Encyclopedia of neuroscience, volume 10. Oxford: Academic
Press, November 2008.
[66] John J. Tukker, W. Rowland Taylor, and Robert G. Smith. Direction selectivity in
a model of the starburst amacrine cell. Visual neuroscience, 21(4):611–625, 2004.
[67] S. Usui, Y . Kamiyama, H. Ish, and H. Ikeno. Reconstruction of retinal horizontal
cell responses by the ionic current model. Vision Research, 36(12):1711–1719,
June 1996.
[68] John P. Vessey, Anna K. Stratis, Bryan A. Daniels, Noel Da Silva, Michael G. Jonz,
Melanie R. Lalonde, William H. Baldridge, and Steven Barnes. Proton-Mediated
Feedback Inhibition of Presynaptic Calcium Channels at the Cone Photoreceptor
Synapse. The Journal of Neuroscience, 25(16):4108–4117, April 2005.
110
[69] Yingxue Wang and Shih-Chii Liu. Motion detection using an aVLSI network of
spiking neurons. In Circuits and Systems (ISCAS), Proceedings of 2010 IEEE
International Symposium, pages 93–96, May 2010.
[70] S. M. Wu. Input-output relations of the feedback synapse between horizontal cells
and cones in the tiger salamander retina. Journal of Neurophysiology, 65(5):1197–
1206, May 1991.
[71] S. M. Wu. Feedback connections and operation of the outer plexiform layer of the
retina. Current opinion in neurobiology., 2(4):462–468, August 1992.
[72] Samuel M. Wu. Synaptic Organization of the Vertebrate Retina: General Principles
and Species-Specific Variations. Investigative Ophthalmology & Visual Science,
51(3):1264–1274, March 2010.
[73] Samuel M. Wu, Fan Gao, and Bruce R. Maple. Functional architecture of synapses
in the inner retina: Segregation of visual signals by stratification of bipolar cell
axon terminals. J. Neurosci., 20(12):4462–4470, June 2000.
[74] D. Xin and S. A. Bloomfield. Dark- and light-induced changes in coupling between
horizontal cells in mammalian retina. The Journal of comparative neurology,
405(1):75–87, March 1999.
[75] T. Yagi. Interaction between the soma and the axon terminal of retinal horizontal
cells in cyprinus carpio. The Journal of physiology, 375:121–135, June 1986.
[76] K. Yoshida, D. Watanabe, H. Ishikane, M. Tachibana, I. Pastan, and S. Nakanishi.
A key role of starburst amacrine cells in originating retinal directional selectivity
and optokinetic eye movement. Neuron, 30(3):771–780, June 2001.
[77] Kareem A. Zaghloul and Kwabena Boahen. A silicon retina that reproduces signals
in the optic nerve. Journal of Neural Engineering, 3(4):257+, December 2006.
[78] Ai-Jun Zhang and Samuel M. Wu. Receptive fields of retinal bipolar cells are
mediated by heterogeneous synaptic circuitry. J. Neurosci., 29(3):789–797, Jan-
uary 2009.
111
Appendix
In this section, we present the transistors’ size, the bias voltages, and the voltage levels
of the power supply for the retinal circuits that we used to perform the simulations in the
previous chapters. We used TSMC 180nm CMOS technology from North Carolina State
University (NCSU) Cadence Design Kit to simulate the circuits. Others may duplicate
our results by utilizing the same technology and the parameters that we provide in this
section.
Figure 5.1 shows the CMOS photoreceptor design. Figure 5.2 shows the glutamate
reuptake design and horizontal cell compartment circuit. Figure 5.3 shows the CMOS
Bipolar Cell Circuit Design. Figure 5.4 shows the CMOS starburst amacrine cell cir-
cuit. Figure 5.5 shows the CMOS wide field amacrine cell circuit. Figure 5.6 and
Figure 5.7 show the sublinear and the single spiking circuit that, together, form a com-
plete object-motion-sensitive retinal ganglion cell. Figure 5.8 shows the Directionally-
Selective Ganglion Cell Circuit. Figure 5.9 shows On bipolar cell and On ganglion Cell
design that we used to demonstrate center-surround property.
112
Figure 5.1: CMOS Photoreceptor Circuit
113
Figure 5.2: Glutamate Reuptake and Horizontal Cell Compartment Circuit
114
Figure 5.3: CMOS Bipolar Cell Circuit
115
Figure 5.4: CMOS Starburst Amacrine Cell Circuit
116
Figure 5.5: CMOS Wide Field Amacrine Cell Circuit
117
Figure 5.6: The Sublinear V oltage Adder in Object-Motion-Sensitive Ganglion Cell
118
Figure 5.7: The Single Spiking Circuit
119
Figure 5.8: The Directionally-Selective Ganglion Cell Circuit
120
Figure 5.9: On Bipolar Cell and On Ganglion Cell for Demonstrating Center-Surround Property
121

Abstract (if available)

Abstract In the biological retina, the feedback and lateral pathways among retinal neurons construct a complicated network that contributes to motion sensing in the retina. When these complex pathways and diverse retinal cell types collaborate, the retina effectively extracts useful information from the visual scene and communicates it to the brain. A silicon retina with motion sensing may be useful for service robots, autonomous vehicles and other applications that require processing dynamic visual information in real time. Implementing motion sensing in a silicon retina presents many challenges. For engineers trying to model the motion sensing functions of a silicon retina, connectivity is one of the most significant engineering challenges that have to be considered. ❧ For this dissertation research, we implement a portion of the starburst amacrine cell (SAC) and differential motion detection model. We also investigate the importance of the feedback and lateral connections in implementing these motion sensing functions in silicon circuit. To validate the importance of the feedback and lateral pathways in the silicon retina, we first build a portion of a retinal network from photoreceptors to ganglion cells that maintains a hierarchical structure similar to that of the biological retina. Lateral connections with horizontal cells and amacrine cells are implemented, along with feedback within the inner and outer plexiform layers of the retina. We then perform demonstrations by comparing the silicon retina tested to one altered by removing these pathways and observing how the behaviors in the silicon retina are changed. We also compare some of our simulation results with biological data. In this research, we showed that some functions cannot be achieved or performances degrade without feedback and lateral connections. Hence, we concluded that incorporating feedback and lateral connections in the artificial retina helps the performance even though it complicates the retinal network.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Plasticity in CMOS neuromorphic circuits

PDF

Dendritic computation and plasticity in neuromorphic circuits

PDF

Adaptive event-driven simulation strategies for accurate and high performance retinal simulation

PDF

Astrocyte-mediated plasticity and repair in CMOS neuromorphic circuits

PDF

Modeling astrocyte-neural interactions in CMOS neuromorphic circuits

PDF

Electrical stimulation of degenerate retina

PDF

Understanding the degenerate retina's response to electrical stimulation: an in vitro approach

PDF

Circuit design with nano electronic devices for biomimetic neuromorphic systems

PDF

Dynamic neuronal encoding in neuromorphic circuits

PDF

Stimulation strategies to improve efficiency and temporal resolution of epiretinal prostheses

PDF

Contributions to structural and functional retinal imaging via Fourier domain optical coherence tomography

PDF

Towards a high resolution retinal implant

PDF

Dependence of rabbit retinal synchrony on visual stimulation parameters

PDF

Motion pattern learning and applications to tracking and detection

PDF

Demonstrating the role of multiple memory mechanisms in learning patterns using neuromorphic circuits

PDF

A variation aware resilient framework for post-silicon delay validation of high performance circuits

PDF

Improving stimulation strategies for epiretinal prostheses

PDF

Cortical and subcortical responses to electrical stimulation of rat retina

PDF

Manipulation of RGCs response using different stimulation strategies for retinal prosthesis

PDF

Intraocular and extraocular cameras for retinal prostheses: effects of foveation by means of visual prosthesis simulation

Asset Metadata

Creator Tseng, Ko-Chung (author)

Core Title Neuromorphic motion sensing circuits in a silicon retina

School Andrew and Erna Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 10/03/2012

Defense Date 07/11/2012

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag motion detection,neuromorphic circuit,OAI-PMH Harvest,silicon retina

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Parker, Alice C. (committee chair), Jenkins, Brian Keith (committee member), Weiland, James D. (committee member)

Creator Email kochungt@usc.edu,kochungtseng@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-99615

Unique identifier UC11289865

Identifier usctheses-c3-99615 (legacy record id)

Legacy Identifier etd-TsengKoChu-1223.pdf

Dmrecord 99615

Document Type Dissertation

Rights Tseng, Ko-Chung

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA