Page 1 |
Save page Remove page | Previous | 1 of 139 | Next |
|
small (250x250 max)
medium (500x500 max)
large ( > 500x500)
Full Resolution
All (PDF)
|
This page
All
Subset |
OPTIMIZING ONLINE LEARNING CAPACITY IN A BIOLOGICALLY-INSPIRED
NEURAL NETWORK
by
Xundong Wu
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
(NEUROSCIENCE)
August 2011
Copyright 2011 Xundong Wu
Object Description
| Title | Optimizing online learning capacity in a biologically-inspired neural network |
| Author | Wu, Xundong |
| Author email | xundongw@usc.edu;wuxundong@gmail.com |
| Degree | Doctor of Philosophy |
| Document type | Dissertation |
| Degree program | Neuroscience |
| School | College of Letters, Arts And Sciences |
| Date defended/completed | 2011-04-29 |
| Date submitted | 2011-07-05 |
| Date approved | 2011-07-06 |
| Restricted until | 2011-07-06 |
| Date published | 2011-07-06 |
| Advisor (committee chair) | Mel, Bartlett |
| Advisor (committee member) |
Zhang, Li Baudry, Michel Sommer, Fritz Singh, Manbir |
| Abstract | To function well in a complex world, our brains must somehow stream our everyday experiences into memory in real time as they occur. An ""episodic"" memory of this kind requires that we form durable memory traces based on single brief exposures to each new piece of incoming information, while preserving old experiences to the greatest extent possible. This combined need for rapid storage and large capacity presents a huge challenge for our nervous system, and we currently know little about how this ""online"" learning process unfolds in detail inside our brains. ❧ Using computer models and mathematical analysis, we have studied the online learning capabilities of a biologically-inspired neural network, with the goal to understand how the properties of neurons, dendrites, and synapses in the network work together to maximize online storage capacity. Based on an accumulation of experimental evidence over the last 20 years, a key assumption in our work has been that dendrites, rather than whole neurons, are the main signaling units used to store and read out learned information. ❧ In the first part of this study, we focused on synaptic plasticity rules and their impact on online storage capacity. We found that online learning was most efficient when: (1) only a tiny fraction of the synapses in the network are modified as each pattern is stored, (2) the learning process -- which involves strengthening synapses -- is confined to a small number of strongly activated dendrites for each pattern, and is gated by both pre- and post-synaptic learning thresholds, and (3) the forgetting process -- which involves depressing synapses -- targets the least recently potentiated synapses within each dendrite undergoing plasticity. ❧ In the second part of this work, given that learning operates at the level of dendrites, we have identified the factors that determine the optimal dendrite size, counted in terms of the number of synapses that are contained within each dendrite. We show that storage capacity is maximized for dendrites of ""medium"" size. Using both simplified probability-based models and large scale simulations, we show why both short and long dendrites suffer from serious capacity costs, including the problem of wasteful over-representation in the case of long dendrites, and in the case of short dendrites, (1) increased susceptibility to noise, (2) increased readout failure rates, and (3) an apparently novel problem arising from the lack of ""dendrite availability"". To increase the applicability of our findings to different brain areas and operating conditions, we also describe the relationship between preferred dendrite size and various characteristics of the input patterns, including (1) pattern density (i.e. the fraction of activated axons), (2) the level of input noise (i.e. the variability in burst strength across activated axons across trials), and (3) the degree of correlation between axons (i.e. the number of axons that have overlapping receptive fields and tend to fire together). We found that noise pushes the optimal dendritic morphology towards longer dendrites, whereas increased density and correlations push the optimal morphology towards short dendrites. ❧ By helping to flesh out the causal chain that links the properties of neurons, dendrites, and synapses to memory capacity, our results can help us understand not only the normal functioning of memory-related areas of the brain, but can also provide conceptual tools needed to understand which types of changes in neurons, their interconnections, and their plasticity rules that occur in aging, neurological disorders, mental retardation, and even stress, are likely to be most detrimental to memory function and why. |
| Keyword | AMPA; NMDA; learning and memory; dendrite; palimpsest memory; online learning; episodic memory; Ubiquitin; brain physiology; computer simulation; dendrites physiology; mathematics; memory physiology; model; neuronal plasticity; neural physiology; synapse physiology |
| Language | English |
| Part of collection | University of Southern California dissertations and theses |
| Publisher (of the original version) | University of Southern California |
| Place of publication (of the original version) | Los Angeles, California |
| Publisher (of the digital version) | University of Southern California. Libraries |
| Provenance | Electronically uploaded by the author |
| Type | texts |
| Legacy record ID | usctheses-m |
| Rights | Wu, Xundong |
| 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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given. |
| Repository name | University of Southern California Digital Library |
| Repository address | USC Digital Library, University of Southern California, University Park Campus MC 7002, 106 University Village, Los Angeles, California 90089-7002, USA |
| Repository email | cisadmin@usc.edu |
| Archival file | uscthesesreloadpub_Volume71/etd-WuXundong-41.pdf |
Description
| Title | Page 1 |
| Full text | OPTIMIZING ONLINE LEARNING CAPACITY IN A BIOLOGICALLY-INSPIRED NEURAL NETWORK by Xundong Wu 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 (NEUROSCIENCE) August 2011 Copyright 2011 Xundong Wu |
Comments
Post a Comment for Page 1
