Thesis (Ph. D.)--University of Rochester. Department of Brain & Cognitive Sciences, 2022.
To develop a complete understanding of visual perception, it is important to investigate the system at all three levels proposed by Marr: computational goal of the visual system, algorithm/representation used to achieve the goal and its implementation in biological circuits. The first part of this thesis explores Marr’s first two levels by trying to understand the consequences and implications of sampling based approximate inference for perceptual decision making and the last part connects it to Marr’s third level by building a network of leaky integrate and fire (LIF) neurons that can perform sampling based inference. Integrating evidence is one of the most elementary computations carried out by the brain, either while passively viewing new information or actively seeking new information. Our previous work showed in simulations and psychohysics experiments, that sampling based approximate inference in a hierarchical model can explain temporal biases arising from biased interpretation of
noisy evidence, during passive viewing of new and independent pieces of information. The first part of this thesis builds on and extends our previous work in three ways: (1) connects the nature of temporal biases to cognitive confidence judgements and tests model predictions for the same on human subjects (2) identifies and manipulates the factors responsible for the temporal biases to generate testable predictions from the
model and verifies them on human subjects using carefully designed psychophysics experiments (3) investigates if human subjects show confirmation bias when seeking new information in a novel gaze contingent evidence integration task and uses sampling based approximate inference to explain the mechanisms of active inference
in decision making. Finally, the last part of this thesis connects all three levels of Marr, by showing that a network of LIF neurons (implementation level) can implement Gibbs sampling based (algorithm/representation level) probabilistic inference (computational level) while manifesting empirical properties of real neurons in the presence of synaptic noise.
