CAREER: Approximate inference at the intersection of neuroscience and machine learning

University Of Rochester, Rochester NY

Investigators

Abstract

Finding patterns in complex data is of great importance to modern society. It lies at the heart of forecasting in business and economics, analyzing data from large-scale experiments, and powering the ongoing revolution in artificial intelligence and machine learning. The brain faces exactly the same challenge: how to extract behaviorally relevant patterns in the large amounts of data it receives from its senses, including millions of photoreceptors in the eyes. This project will investigate two important aspects of the computations underlying this process: how to decide whether to combine two pieces of information, for instance, from two photoreceptors, or even two different senses, and the consequences of doing so not instantaneously and exactly, but over time and approximately. By investigating these questions in the brain, this project aims to extract insights that are relevant for both neuroscience and machine learning. The main contributions of this project to neuroscience and cognitive science are a deeper understanding of the central computational motif underlying sensory processing and new computational theories of confirmation bias and attention. The central contributions to machine learning will consist in suggesting architectural changes to current deep learning architectures and in evaluating their performance benefits. In its educational part, this project will develop and evaluate a curriculum for an interdisciplinary, research project-based college-level course for educating the next generation of computational neuroscience and machine learning researchers. This project focuses on two key questions: (1) What is the architecture of the deep probabilistic model that the brain has learned, and (2) How does the brain perform approximate and sequential, as opposed to exact and instantaneous, inference in this model? The researchers will address the first question by proposing a deep hierarchical causal inference-based model for motion perception. They will quantitatively test this model using human psychophysical experiments and collaborate to test it using neurophysiology experiments. The relationship of the probabilistic model's central computational motif to other computations such as divisive normalization and predictive coding will be explored. In collaboration with machine learning researchers the benefits of incorporating the computational insights into deep learning systems will be quantified. The second research aim investigates the consequences of approximate inference computations in two contexts. First, the researchers will develop a rigorous computational theory relating limited biological and computational "resources" during approximate inference. Second, they will compare the biases due to approximate inference with those found in humans using psychophysical evidence integration tasks. In preliminary data the emergence of a confirmation bias as the result of approximate inference was found -- both in the context of passive interpretation of visual evidence, and in the context of active inference using eye-movements. In addition to a better understanding of inference in the brain, the goal of this work is to yield insights into strategies for how to counter biases and design efficient artificial intelligence systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.