Ralf Haefner, Ph.D.

Associate Professor

School of Arts and Sciences: Brain and Cognitive Sciences

University of Rochester

Behavioral and neural signatures of generative inference in the brain

It has long been understood that perception is an inference process during which the brain combines prior experience with incoming sensory information. Two influential frameworks guide the field’s thinking about these inference computations and their implementation in neural and artificial systems. In one, sensory processing is a largely feedforward, discriminative process that transforms sensory observations into useful representations. In another, the brain inverts a generative model by combining top-down predictions with bottom-up inputs. In my talk I will present behavioral and neurophysiological evidence for the second view, sometimes called “generative inference”. Behaviorally, I will show how approximate generative inference correctly predicts the strength of confirmation bias and associated overconfidence during evidence accumulation. Neurophysiologically, I will show that the same model correctly predicts the task-dependence of correlated variability (noise correlations) between sensory neurons in monkey V1. Interestingly, the predicted noise correlations imply that the redundancy of sensory responses should increase with task learning – in contradiction to classic theories of learning based on redundancy reduction. By tracking population responses in macaque cortical area V4 as monkeys learned visual discrimination tasks we found strong evidence for this generative inference prediction.

Faculty Host: Dr. Chris Fetsch