Title: Biologically Inspired Learning for Perception, Exploration, and Decision-Making Under Uncertainty

Date: October 7, 2026

Time: 1-3 PM ET

Location: Price Gilbert 4222

Zoom link:

https://gatech.zoom.us/j/98261111313?pwd=xtWSnKAbZJ0ya4ui86blMrJ76ZOqfM.1

Zoom passcode: 599523

Abdelrahman Sharafeldin

Machine Learning PhD Student

The Wallace H. Coulter Department of Biomedical Engineering

Georgia Institute of Technology

Committee

1 Dr. Hannah Choi (Advisor), School of Mathematics, Georgia Institute of Technology

2 Dr. Nabil Imam, School of Computational Science and Engineering, Georgia Institute of Technology

3 Dr. Anqi Wu, School of Computational Science and Engineering, Georgia Institute of Technology

4 Dr. Simon Sponberg, School of Physics, Georgia Institute of Technology

5 Dr. Chethan Pandarinath, Department of Biomedical Engineering, Georgia Institute of Technology

Abstract

Intelligent behavior requires organisms to learn useful representations of the world, seek informative observations, and act effectively despite uncertainty and limited resources. This thesis investigates how predictive generative perception, information-seeking action, and reward-guided learning support these abilities under biological constraints such as embodiment, energy efficiency, and neural architecture. In the first part of this thesis, we develop an active-sensing framework that uses expected reductions in the uncertainty of a generative model to guide exploration, learning environmental structure and visual representations that improve data efficiency in subsequent tasks. In the second part, we extend this framework to multiple sensory modalities and show that the learned visual and mechanosensory information landscapes account for features of hawkmoth flower probing and tracking. We also develop a mode-switching control algorithm that uses perceptual uncertainty to coordinate exploration and reward seeking and improves learning and generalization in simulated nectar search. In the third part, we ask how these computational objectives for perception and action can be instantiated in cortical circuits and develop a cell-type-specific model that combines predictive coding, reinforcement learning, and activity costs under biological connectivity constraints. Trained on a naturalistic change-detection task, this model reproduces several experimentally observed responses to novelty across different cell types without fitting neural response data. These complementary studies connect how information is acquired, represented, and used, offering insights into biological computation and principles for designing artificial agents that learn through interaction.