How does human vision make good perceptual guesses about objects?

Daniel Kersten

Talk Summary

The speaker provided an explanation of how human vision makes good perceptual guesses about objects using bayesian influence graphs. To this end, he decomposed an object $S$ and its image $I$ into world/object properties and features.The resulting probability distribution was then represented using a graph where nodes represent random variables and links represent the influences. He then illustrated this approach using examples of discounting and cue integration. In the second half of his talk the speaker concentrated on V1 and LOC mechanisms involved in perception. His experiments revealed that perceptual organization correlated with reduced V1 activity and increases LOC activity. He showed that V1 activity can predict percept on the time-scale of behaviour. The decrease in V1 activity could mean two things 1) Predictive Coding 2) Sparsification. In predictive Coding high-level object models project back predictions of the incoming data. In this case a good fit implies a low activity at the lower areas due to subtraction. (The "shut-up" theory). In sparsification, a good high-level fit tells the lower areas to 'stop gossiping". This essentially amplifies the activity for features belong to the object and suppress the rest. Since predictive coding and sparsification have the same empirical fMRI observation, the experiments were inconclusive in deciding which of the two mechanisms (shut-up or stop gossiping) reduces the V1 activity when a higher level (LOC) has a good explanation for an object.

Questions During the Presentation:

• Bill : Could you explain the term Discounting?
• Ans : Discounting is throwing away information that is not required for perception. For eg: lighting

• Jeff: (Regarding shape cues explaining away pinkishness as white paper and pink light)Are these things learnt?
• Ans : Don't know yet. We are in the process of esigning and experiment to find that out.

Discussion after the Presentation:

• Grossberg: In non-stationary environments priors don't exist. Also the bayesian framework breaks apart for some Kanisza illusions.

• Bruno: Nothing in the Bayesian framework is inconsistent with Kanizsa. Bayesian framework is taking out the Kanizsa illusions and explaining those. The correct choice of priors is the trick.

• Dan: For eg: one prior for natural vision is that motion tends to be slow.

• Bill Softky: You showed us data showing suppression of activity in V1. Is there any data for suppression of thalamus?

• Dan: In an fMRI experiment, localizing the thalamus and especially the LGN is not easy. Hence we don't have any data for that.

• This was followed by Steve explaining a Kanizsa illusion.

Back to the main index for Inference and Prediction in Neocortical Circuits.