Adaptation in the Neural Code of the Retina

Transcription

1 Adaptation in the Neural Code of the Retina Lens Retina Fovea Optic Nerve

2 Optic Nerve Bottleneck Neurons Information Receptors: % Optic Nerve 106 5% After Polyak 1941 Visual Cortex ~1010

3 Mean Intensity Varies a Lot, But Slowly

4 Light Adaptation: Evolutionary Rods and cones have 1000x different sensitivity to light Rod Cone Schneeweis & Schnapf 1999, 2000

5 Light Adaptation: Dynamic Sensitivity decreases with increasing background light Rod Cone Schneeweis & Schnapf 1999, 2000

6 Ganglion Cells Report Intensity Ratios Response vs intensity for cat retinal ganglion cells adapted to increasing backgrounds. Sakmann & Creutzfeld 1969

7 Contrast Varies, But Slowly Intensity Distribution Log (Intensity) Retinal Sensitivity Log (Intensity)

8 Response sensitivity adapts to stimulus contrast Contrast Time (s)

9 Contrast Adaptation Occurs over Seconds I = C = Firing Rate (Hz) τ = 14.9 s τ = 26.3 s Time (s) 200 Smirnakis et al 1997

10 Probing the Origins of Contrast Adaptation A B - + G

11 Natural Scenes Are Highly Correlated in Space Joint probability, PI (, I ) 1 2 Intensity of Neighboring Pixel 2 Intensity of Pixel 1

12 Natural Scenes Have 1/k 2 Spatial Power Spectra Amplitude Spectrum Field 1987

13 Ganglion Cells Report Spatial Differences Cat ganglion cell receptive field probed with a thin white bar 50 Hz 10 deg Bar position Rodieck & Stone 1965

14 Natural Scenes are Correlated in Time t t + 33 ms Joint Probability Intensity at t+33ms Intensity at t increases decreases Dong & Atick 1995

15 Ganglion Cells Report Temporal Differences Macaque ganglion cell response kernels ON Cell OFF Cell Time from spike (s) Time from spike (s) Chichilnisky & Kalmar 2002

16 Cone Photoreceptors Compute Temporal Differences Macaque cone Schnapf et al 1990

17 Predictive Coding 1 Response I0 ( I-1+ I1) " 2 $ # $ % Prediction for I0 from adjacent points " $$$$ # $$$$ % Deviation of I0 from prediction 1 Response I I- + I 2 ( ) Suppresses the predominant, predictable patterns. Enhances unexpected, novel features: e.g. edges, motion. Thought to be an evolutionary adaptation. Attneave, Barlow, Srinivasan, Laughlin, Atick, et al

18 Dynamic Adaptation to a Change in Correlations?

19 Probing Adaptation to Correlations Adapt Probe Adapt Probe 10 s 1.5 s 10 s 1.5 s Ganglion cell sensitivity to the two stimuli:

20 Oriented Patterns Modify Receptive Fields After adaptation to: Sensitivity to:

21 Many Ganglion Cells Adapt to Oriented Patterns After adaptation to: Ratio αh Ratio αv Adaptation Index: A = α H α V Sensitivity to: Cells A A

22 Receptive Fields Adapt to Spatial Frequency Adapting Stimuli: High Frequency Low Frequency Ganglion cell sensitivity to: Adaptation Index: A

23 Pattern Adaptation Occurs Over Seconds µm 272 µm Firing Rate (Hz) Time (s) 120 Smirnakis et al 1997

24 Filter Functions Adapt to Temporal Correlations Adapting Flicker Stimuli: Adaptation Index: A

25 Responses Adapt to Space-Time Correlations Adapting Stimuli: Ganglion cell sensitivity: Adaptation Index: A

26 Parallel Channels Hypothesis Adapting Stimulus Interneurons Adaptive Gain Ganglion Cell But... Interneurons are not very selective, e.g. bipolar cells: Requires interneurons selective for many kinds of patterns, spatial, temporal, spatio-temporal...

27 Adaptive Network Hypothesis Connections are plastic using anti-hebbian rule: If AC and GC strongly correlated, then inhibition strengthens. ACs try to predict GC signal based on signals at other points in space. Successful predictions get subtracted. Other forms of prediction across space and time, using amacrine cells of diverse types: small, large, transient, sustained,...

28 Anti-Hebbian Synaptic Plasticity Linear processing: y = ( b + a ) x Synaptic plasticity: i ij ij j j d dt a 1 ij = a ij β y i x j τ, β > τ 0 Decay term Anti-Hebbian term: More inhibition when pre and postsynaptic neurons are correlated. Foldiak, Barlow

29 Model Adaptation to Grating Stimuli Ganglion cell starts with non-oriented receptive field: Adapting stimulus: Receptive field:

30 Adaptive Network Hypothesis

31 Blockers of Inhibition Reduce Pattern Adaptation Control Picrotoxin + Strychnine Adaptation to: Sensitivity to: Adaptation Index: A A

32 Probing the Inhibitory Strength of Amacrine Cells 1 0 I (na) Time From Amacrine Pulse (s) Hz

33 Amacrine-to-Ganglion Cell Inhibition is Modulated by Pattern Adaptation A-to-G Correlation During Adaptation Horizontal Vertical A-to-G Inhibition After Adaptation Horizontal Vertical G1: Delay (s) G2: Delay (s)

34 Conclusions Much of retinal adaptation can be understood as suppressing predictable signals. The rules used for prediction adjust dynamically to a change in the correlation structure of the visual input. This adaptation serves to emphasize novel features. It can make a substantial contribution to various pattern adaptations in human perception. Dynamic predictive coding could be implemented through anti-hebbian synaptic plasticity at inhibitory synapses.