Session 5: Miguel Eckstein, PhD, Distinguished Professor, University of California, Santa Barbara

Title: From Bayesian ideal observers to deep neural networks

For four decades, Bayesian ideal observer models (BIO) have been an important tool to understand human vision. They have been used as benchmarks to compare human perceptual performance against the upper optimal bound, assess whether behaviors arise as byproducts of task information or stimuli priors, and identify sources of suboptimalities in human visual processing. The main limitation of the BIO is that its computation requires full knowledge of the image statistics and cannot be applied to real world scenes without strong feature extraction assumptions. Thus, this limits the use of the BIO to understand natural tasks.

In contrast, Deep Neural Networks (DNNs) can be applied to any image set including real world scenes. However, they do not guarantee optimality and their computational stages are harder to interpret. In this talk, I will discuss how we can learn about properties of human vision with real world scenes and tasks from comparisons of DNNs and human behaviors and from comparisons of the inner properties of DNNs and BIO models.

Видео Session 5: Miguel Eckstein, PhD, Distinguished Professor, University of California, Santa Barbara канала The Smith-Kettlewell Eye Research Institute