how recurrent dynamics explain crowding

1
How Recurrent Dynamics Explain Crowding Aaron Clarke & Michael H. Herzog Laboratory of Psychophysics, Brain Mind Institute, École Polytechnique Fédérale de Lausanne (EPFL), Switzerland Introduction: Crowding is the inability to discriminate objects in clutter Vernier discrimination, for example, deteriorates when the Vernier is flanked by parallel lines Pooling (Wilkinson et al., 1997) and lateral inhibition (Wilson, 1986) models predict that adding more parallel lines should worsen performance S S Lateral Inhibit ion: Spatia l Poolin g: Inhibit ory Excitato ry Important points about the human data: 1. Effects of adding more lines depends critically on line-length Information for lines of different lengths flows through separate channels and may be combined based on a length-based similarity metric To model this we implement an end-stopped receptive field filter-bank sensitive to lines of different lengths Connection strength between filters selective for different line-lengths depends on their similarity W Btw W Btw 2. Parallel lines of the same length cause maximal interference Lateral interactions between parallel receptive fields modulate the cells’ outputs This may be modeled by weighting connections between cells with parallel receptive fields using a lateral-inhibitory association field Figure 1. Human data. Performance worsens when equal-length flankers are added, but improves when shorter- or longer-length flankers are added (Malania et al., 2007). 2 16 10 20 30 40 50 60 Flanks (#) Threshold (arcsec) Human Data Adding more lines can, however, improve performance We propose that performance worsens when the flankers group with the Vernier, but improves when the flankers segregate from the Vernier A recurrent architecture employing a Wilson- Cowan type model can explain these results because it allows local information to propagate globally over time Global grouping arises from local, dynamical interactions without explicit grouping rules Conclusions: Crowding cannot be explained by lateral inhibition or spatial pooling models Crowding can be explained by a Wilson-Cowan type model Global grouping arises through local dynamics without explicit grouping rules Redundant information is suppressed while inhomogeneities are highlighted References: • Malania, M., Herzog, M.H. &Westheimer, G. (2007). Grouping of contextual elements that affect Vernier thresholds. Journal of Vision. 7(2):1, 1-7. • Wilkinson, F., Wilson, H.R. & Ellemberg, D. (1997). Lateral interactions in peripherally viewed texture arrays. J. Opt. Soc. Am. A. 14(9): 2057-2068. • Wilson, H.R. (1986). Responses of Spatial Mechanisms Can Explain Hyperacuity. Vision Research. 26(3):453-469. • Wilson, H.R. & Cowan, J.D. (1972). Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons. Biophysical Journal. 12:1-24. http://lpsy.epfl.ch This work was supported by the ProDoc project “Crowds in Crowding" of the Swiss National Science Foundation (SNF) Corresponding author: [email protected] Model Specifics: E I E I E I Excitatory Inhibitor y E I E I E I E I E I E I Excitat ory Layer Inhibito ry Layer End-stopped receptive field array •X Input Image Linking Hypothesis: 0.75 0.8 0.85 0.9 0.95 1 15 20 25 30 35 40 45 Cross-Correlation With Vernier Template Vernier Threshold (arc sec) Data Fit •X •X 2 16 10 20 30 40 50 60 Flanks (#) Threshold (arcsec) Model Data No flanks Short Equal Long Figure 2. In the end the summed cross- correlations are passed through a sigmoidal non- linearity. Figure 3. The model nicely predicts the pattern of results obtained by Malania et al. (2007). Cross-correlate un- flanked Vernier template with the flanked Vernier images Sum the cross-correlation outputs over space and filter sizes The model suppresses homogeneities while highlighting inhomogeneities Model outputs for each image at each filter size are cross-correlated (.x) with the outputs for the un-flanked Vernier and summed over filter sizes The equal-length flankers outputs correlate poorly with the un-flanked Vernier outputs (e.g. compare black outlined image with green outlined image) The long-length flankers outputs correlate well with the un-flanked Vernier outputs (e.g. compare black outlined image with blue outlined image) W Btw

Upload: mervin

Post on 24-Feb-2016

30 views

Category:

Documents


2 download

DESCRIPTION

How Recurrent Dynamics Explain Crowding . Aaron Clarke & Michael H. Herzog Laboratory of Psychophysics, Brain Mind Institute, École Polytechnique Fédérale de Lausanne (EPFL), Switzerland. Introduction: Crowding is the inability to discriminate objects in clutter - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: How Recurrent Dynamics Explain Crowding

How Recurrent Dynamics Explain Crowding Aaron Clarke & Michael H. Herzog

Laboratory of Psychophysics, Brain Mind Institute, École Polytechnique Fédérale de Lausanne (EPFL), Switzerland

Introduction:

• Crowding is the inability to discriminate objects in clutter• Vernier discrimination, for example, deteriorates when the Vernier is

flanked by parallel lines• Pooling (Wilkinson et al., 1997) and lateral inhibition (Wilson, 1986)

models predict that adding more parallel lines should worsen performance

S S

Lateral Inhibition:

Spatial Pooling:

Inhibitory

Excitatory

Important points about the human data:

1. Effects of adding more lines depends critically on line-length• Information for lines of different lengths flows through separate

channels and may be combined based on a length-based similarity metric

• To model this we implement an end-stopped receptive field filter-bank sensitive to lines of different lengths

• Connection strength between filters selective for different line-lengths depends on their similarity

WBtwWBtw

2. Parallel lines of the same length cause maximal interference• Lateral interactions between parallel receptive

fields modulate the cells’ outputs• This may be modeled by weighting connections

between cells with parallel receptive fields using a lateral-inhibitory association field

Figure 1. Human data. Performance worsens when equal-length flankers are added, but improves when shorter- or longer-length flankers are added (Malania et al., 2007).

2 1610

20

30

40

50

60

Flanks (#)

Thre

shol

d (a

rcse

c)

Human Data

• Adding more lines can, however, improve performance

• We propose that performance worsens when the flankers group with the Vernier, but improves when the flankers segregate from the Vernier

• A recurrent architecture employing a Wilson-Cowan type model can explain these results because it allows local information to propagate globally over time

• Global grouping arises from local, dynamical interactions without explicit grouping rules

Conclusions:• Crowding cannot be explained by lateral inhibition or spatial pooling

models• Crowding can be explained by a Wilson-Cowan type model• Global grouping arises through local dynamics without explicit grouping

rules• Redundant information is suppressed while inhomogeneities are

highlighted

References:• Malania, M., Herzog, M.H. &Westheimer, G. (2007). Grouping of contextual elements that affect Vernier thresholds. Journal of Vision. 7(2):1, 1-7. • Wilkinson, F., Wilson, H.R. & Ellemberg, D. (1997). Lateral interactions in peripherally viewed texture arrays. J. Opt. Soc. Am. A. 14(9): 2057-2068.• Wilson, H.R. (1986). Responses of Spatial Mechanisms Can Explain Hyperacuity. Vision Research. 26(3):453-469.• Wilson, H.R. & Cowan, J.D. (1972). Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons. Biophysical Journal. 12:1-24.

http://lpsy.epfl.ch This work was supported by the ProDoc project “Crowds in Crowding" of the Swiss National Science Foundation (SNF) Corresponding author: [email protected]

Model Specifics:

⊗⊗

⊗E

I

E

I

E

I

Excitatory

Inhibitory

E

I

E

I

E

I

E

I

E

I

E

I

Excitatory Layer

Inhibitory Layer

End-stopped receptive field array

•X

Input Image

Linking Hypothesis:

0.75 0.8 0.85 0.9 0.95 1

15

20

25

30

35

40

45

Cross-Correlation With Vernier Template

Vern

ier T

hres

hold

(arc

sec

)

DataFit

•X •X

2 1610

20

30

40

50

60

Flanks (#)

Thre

shol

d (a

rcse

c)

Model Data

No flanksShortEqualLong

Figure 2. In the end the summed cross-correlations are passed through a sigmoidal non-linearity.

Figure 3. The model nicely predicts the pattern of results obtained by Malania et al. (2007).

Cross-correlate un-flanked Vernier template with the flanked Vernier images

Sum the cross-correlation outputs over space and filter sizes

• The model suppresses homogeneities while highlighting inhomogeneities

• Model outputs for each image at each filter size are cross-correlated (.x) with the outputs for the un-flanked Vernier and summed over filter sizes

• The equal-length flankers outputs correlate poorly with the un-flanked Vernier outputs (e.g. compare black outlined image with green outlined image)

• The long-length flankers outputs correlate well with the un-flanked Vernier outputs (e.g. compare black outlined image with blue outlined image)

WBtw