题目：A probabilistic approach to visual perception and learning.

报告人：Prof. József Fiser, Central European University

时间：2015年04月28日 星期二 13:00-14:30

地点： 北京大学王克帧楼1113室

In this talk, I will investigate the issue of learning in vision through behavioral studies. First, I will argue that the suitable conceptual framework for studying visual learning is defined as representational learning thatunifies the fragmented areas of perceptual, statistical and rule learning, and that the proper computational framework to investigate representational learning is probabilistic. To support this argument, I will present experimental results showing how even the lowest level of classical perceptual processes are influenced by internal representational factors. Next,
I will briefly introduceand review the earlier results of statistical learning (SL), the type of learning that is closest in spirit to the idea of representational learning. I will focus on the question of where to go next now, after the basic results of SL have been established. I will present new results showing how extending the framework of classical SL can give explanation to a number of discrepancies in published SL results and how this extension justifies the probabilistic approach to learning. Time permitting, I will put these issues into the larger context of perception and learning in the neural substrate of the cortex. I will introduce a sampling-based probabilistic framework of how the cortex mightencode an learn from internal stimuli, and I will provide human behavioral evidence that supports the idea of the cortex and sampling based probabilistic inference and learning machine.