From visual perception to decision: utilizing machine learning to probe encoding and decoding in V1
Edgar Y. Walker, Ph.D.
Postdoctoral Fellow, University of Tübingen:
Seminar abstract: A central goal of visual neuroscience lies in understanding how the visual system allows us to utilize visual sensory information to guide our decision-making and behavior. We can view the process as consisting of two complementary components: encoding and decoding. The visual system encodes information about the visual world into the activities of the visual cortical populations. Subsequently, the rest of the brain decodes the visual information from the population activity and performs computations to arrive at a decision and behavior. Recent advances in recording techniques allow us to collect sensory population data with ever-increasing size and complexity. Such data, especially in combination with rich naturalistic stimuli and carefully designed behavioral tasks, holds great promise in tackling the challenging questions on both encoding and decoding fronts. However, the data’s unprecedented complexity poses a significant challenge to the more conventional population analysis techniques and calls for novel approaches to take full advantage of the data. In this talk, I will present my recent work in which I specifically applied advanced deep learning-based analyses on rich population recordings from V1 in macaques and mice to advance understanding of both encoding and decoding in visual cortical populations. On the encoding front, I developed a state-of-the-art deep neural network (DNN) model of mouse V1 responses to natural images and utilized this network to generate the most exciting image (MEI) for each neuron. The MEIs significantly deviated from the conventional linear receptive fields and Gabor-filters, and their effectiveness in driving target neurons were experimentally verified in a closed-loop experimental paradigm that we called the “inception loop”. On the decoding front, I developed a novel DNN-based technique to decode trial-by-trial uncertainty information about the sensory stimulus from macaque V1 as the monkey performed a task requiring the use of the uncertainty information. Utilizing the decoded uncertainty, I was able to critically assess competing models of uncertainty information propagation in the visual system during decision-making and provided the first population electrophysiological evidence supporting the theory of probabilistic population code, a leading theory on probabilistic computations in the brain.