I will discuss three ideas for investigating deep nets that I want to pursue during my postdoc:
(i) Understanding human rapid visual perception by presenting manipulated versions of objects that need to be detected, and trying to understand which features are critical. (ii) Training deep nets to match human performance on all these tasks. (iii) Using summary statistics in deep nets to perform segmentation.
I need feedback which of these ideas are most interesting and how to pursue them best.