Computational modeling and utilization of attention, surprise and attention gating [slides]

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Explaining Observer Performance in Dynamic Vision Tasks 
using Bayesian Surprise 
 There are many models of feature based attention but in 
general they lack a temporal component to detect unique 
feature changes across time.

Treisman, A.M., & Gelade, G. (1980). Cognitive Psychology, 12 (1), 97‐136. 

Koch, C., & Ullman, S. (1985). Human Neurobiology, 4 (4), 219‐227. 

Itti, L., Koch, C., & Niebur, E. (1998). IEEE PAMI 20 (11), 1254‐1259. 
See also:
Shiffrin, R.M., & Schneider, W. (1977). Psychological Review, 84 (2), 127‐190. 
http://www.nerd‐cam.com 
Koch and Ullman Itti and Koch 
 2AFC Task. Did you see an Animal? (or Transportation 
Method) – Two separate experiments. 
 20 Hz exposure. Natural scene distracters. 
 An attention gate may predict what is attended to in a 
dynamic scene. 
 Can we create model of an attentional gate that extends 
the notion of feature integration into time?

Sperling, G., & Weichselgartner, E. (1995). Psychological Review, 102 (3), 503‐532. 

Shih, S.‐I., & Sperling, G. (2002). Psychological Review, 109 (2), 260‐305. 
 We can predict, using Bayesian 
Surprise, the activity of an attention 
gate.
 The attention gate is a triage system for 
controlling what visual information from 
targets is able to pass initial processing 
to higher visual centers. 
 A triage system for visual information 
assumes an information bottleneck 
which necessitates a choice of 
information selection since not all 
information can be processed by 
increasingly complex visual systems.
 Hypothesis: A large part of RSVP performance can be 
explained in terms of an attentional gate (Sperling et 
al, Cave, Chun & Potter, etc…)
 What gets past the attention gate is perceivable. 
 Things which are more interesting or more important; 
perhaps more informative, should have a better ability 
to control the attention gate. 
 The attention gate triages information from RSVP. 
 Can we create and test such a mechanism? 

Sperling, G., & Weichselgartner, E. (1995). Psychological Review, 102 (3), 503‐532. 

Cave, K.R. (1999). Psychological Research, 62, 182‐194. 

Chun, M.M., & Potter, M.C. (1995). Journal of Exp. Psychology: Human Perc. and Perf., 21, 109‐127. 
 Given an input stream of images, 
what is truly new and informative? 
 Information outliers are surprising. 
 We should be able to resist garbage 
information like 1/f noise.
 Information is based on image 
features (Treisman & Gelade, Koch 
& Ullman, Itti & Koch). 
 What is informative should be 
better able to pass the Attention 
Gate. 

Treisman, A.M., & Gelade, G. (1980). Cognitive Psychology, 12 (1), 97‐136. 

Koch, C., & Ullman, S. (1985). Human Neurobiology, 4 (4), 219‐227. 

Itti, L., Koch, C., & Niebur, E. (1998). IEEE PAMI 20 (11), 1254‐1259. 
Blue/Yellow Color Opponent 
 Shannon (1948):
D = dataset D = all possible
dataset
 Problems:
 Fine for communication; but what about
semantic/subjective aspects?
 Information vs. value, importance, relevance, or
surprise.
 I is informative compared with what?
 White snow paradox.
TV news, sports, music,
action movies, etc
0.3 MByte/s
(640x480, MPEG4,
46,000 frames)
Greyscale snow
5.0 MByte/s
Shift emphasis:
 from objective probability of occurrence of data
 …to effects of data onto subjective beliefs of observers.
. . .
P(M)
M
MTV CNN FOX BBC . . . Snow
prior
 Family M of observer-dependent models or hypotheses
about the world.
 Observer beliefs:
 Bayesian foundation of probability: data is what changes a
prior into a posterior:
. . .
P(M)
M
MTV CNN FOX BBC . . . Snow
prior
. . .
P(M)
M
MTV CNN FOX BBC . . .
Snow
prior
. . .
P(M|D)
M
MTV CNN FOX BBC . . .
Snow
posterior
. . .
P(M)
M
MTV CNN FOX BBC . . .
Snow
prior
. . .
P(M|D)
M
MTV CNN FOX BBC . . . Snow
posterior
Surprise
. . .
M
MTV CNN FOX BBC . . . Snow
Surprise =
P(M),
P(M|D)

Beliefs stabilize, prior and posterior become identical,
and additional snow frames carry no surprise.
. . .
M
MTV CNN FOX BBC . . . Snow
P(M),
P(M|D)

 Surprise:
 using, e.g., the Kullback-Leibler (KL) distance for d.
 Shannon’s Information:
 Moral: We want relative information rather than
absolute information.

We want to start to quantify surprise by models of 
something directly measurable such as image features. 
 This is an easy way to quantify an image. 

Models of image features are the expected feature response 
given past feature measurements. 

Approximate P(M) and P(M|D) with a probability 
distribution. 
Feature Response ‐ Models 
Use a Gamma probability over feature 
responses since we assume Poisson 
noise (Neural Spike Trains) and a 
response range from zero to infinity
So… How can Updating our Beliefs Surprise us? 
 Create models over all features and locations in images 
then combine into Surprise Maps. 
 Recall the RSVP task you saw earlier. Can we make 
predictions about how observers will perform on it? 
 Run 8 subjects on 500 RSVP sequences with Animal 
Targets and natural distracters (see Example Below) 
 Some are  and 
subjects always spot 
the target. 
 Some are  and 
subjects always miss 
the target. 
 Can surprise tell use 
why? 
 Can we make  RSVP sequences  based on surprise 
statistics? 
 Idea: Change order of images to block the target with images 
which are more Surprising than it. 
Einhäuser, W., Mundhenk, T.N., Baldi, P., Koch, C., & Itti, L. (2007). Journal of Vision, 7 (10), 1‐13. 
 Can we make “Easy” RSVP sequences hard based on 
surprise statistics? 
Einhäuser, W., Mundhenk, T.N., Baldi, P., Koch, C., & Itti, L. (2007). Journal of Vision, 7 (10), 1‐13. 
 The M‐W Pattern Emerges 
 +/‐ 50 MS is critical for 
Surprise masking in RSVP 
 W – Easy 
 M – Hard 
Mundhenk, T.N., Einhäuser, W., & Itti, L. (2009). Vision Research, In Press 
Mundhenk, T.N., Einhäuser, W., & Itti, L. (2009). Vision Research, In Press 
Mundhenk, T.N., Einhäuser, W., & Itti, L. (2009). Vision Research, In Press 
For Example Chun and Potter 1995
 Can we recreate the attention gate to reveal what 
image contents are detectable? 
 Easy target locations should have a higher likelihood 
of passing the attention gate than hard target 
locations.  
 Build on the idea that more surprise for a target 
means more pass through. 
 Note: We expand the data set to include 
transportation targets in addition to the animal 
targets. 
 If the attention gate is valid, then the surprise attention gate 
should overlap more with easy targets than with hard targets. 
 Thus, targets are easier because more image information gets 
past the attention gate. 

 Surprise seems to reveal the activity of the attentional 
gate for quick automatic attention. 
 Surprise at the feature level supports a two‐stage model 
of attention and agrees with data on lag sparing and 
attentional blink in RSVP. 
 We have greater insight into what parts of a dynamic 
scene can be perceived by observers. 
 The current Attention Gate model of Bayesian Surprise works 
very well with different types of targets. 
 The model does need some more work and testing. 
 The model has some neat properties such as giving an 
explanation to split attention effects. 
 Surprise may unify saliency maps with automatic attention 
gating. 
• Itti, L., & Baldi, P. (2006). Bayesian Surprise attracts human attention. Advances in Neural Information 
Processing Systems (NIPS), 19 (pp. 547‐554): MIT Press. 
• Einhäuser, W., Mundhenk, T.N., Baldi, P., Koch, C., & Itti, L. (2007). A bottom‐up model of spatial 
attention predicts human error patterns in rapid scene recognition. Journal of Vision, 7 (10), 1‐13. 
• Mundhenk, T.N., Einhäuser, W., & Itti, L. (2009). Automatic Computation of an Image’s Statistical 
Surprise Predicts Performance of Human Observers on a Natural Image Detection Task. Vision Research, 
In Press 
• http://www.mundhenk.com/thesis/

Asset Metadata

Creator Mundhenk, Terrell Nathan (author)

Core Title Computational modeling and utilization of attention, surprise and attention gating [slides]

Contributor Electronically uploaded by the author (provenance)

School Andrew and Erna Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Degree Conferral Date 2009-08

Publication Date 08/04/2009

Defense Date 04/21/2009

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

Tag attention,attentional blink,bayes,biologically inspired,CINNIC,computation,contour,detection,gating,H2SV,Human Performance,iLab,image processing,information,Integration,Itti,Koch,MAG,masking,Nerd-Cam,Neuromorphic Vision Toolkit,OAI-PMH Harvest,RSVP,saliency,spot light,statistics,Surprise,tracking,vision,visual cortex,visual saliency

Format 36 pages (extent)

Language English

Advisor Itti, Laurent (committee chair), Arbib, Michael A. (committee member), Biederman, Irving (committee member), Schaal, Stefan (committee member)

Creator Email nathan@mundhenk.com,tnmundhenk@hrl.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c127-15525

Unique identifier UC188659

Identifier usctheses-c127-15525 (legacy record id)

Legacy Identifier etd-Mundhenk-2997-defense_slides

Dmrecord 15525

Document Type Dissertation

Format 36 pages (extent)

Rights Mundhenk, Terrell Nathan

Internet Media Type application/pdf

Type texts

Source University of Southern California (contributing entity)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email uscdl@usc.edu

Abstract (if available)

Abstract What draws in human attention and can we create computational models of it which work the same way? Here we explore this question with several attentional models and applications of them. They are each designed to address a missing fundamental function of attention from the original saliency model designed by Itti and Koch. These include temporal based attention and attention from non-classical feature interactions. Additionally, attention is utilized in an applied setting for the purposes of video tracking. Attention for non-classical feature interactions is handled by a model called CINNIC. It faithfully implements a model of contour integration in visual cortex. It is able to integrate illusory contours of unconnected elements such that the contours â€œpop-outâ€ as they are supposed to and matches in behavior the performance of human observers. Temporal attention is discussed in the context of an implementation and extensions to a model of surprise. We show that surprise predicts well subject performance on natural image Rapid Serial Vision Presentation (RSVP) and gives us a good idea of how an attention gate works in the human visual cortex. The attention gate derived from surprise also gives us a good idea of how visual information is passed to further processing in later stages of the human brain. It is also discussed how to extend the model of surprise using a Metric of Attention Gating (MAG) as a baseline for model performance. This allows us to find different model components and parameters which better explain the attentional blink in RSVP.