2025 Theses Doctoral

Modular Learning Systems for Continual Control: Neural Principles and Computational Models

Amematsro, Elom Andoche

How do animals learn flexible behaviors that generalize across time, context, and perturbation? This thesis addresses this question through a unified framework that links normative theories of control, neural population dynamics, and learning algorithms.

Chapter 2: A Framework for Motor Control introduces a probabilistic framework for motorcontrol that unifies principles from optimal feedback control and dynamical systems theory. By casting control as inference in a latent dynamical system, I show that internal memory dynamics and sensory feedback jointly support adaptive, feedback-sensitive motor behavior. The resulting model naturally reproduces hallmark features of biological control—including preparatory activity, feedback corrections, and orthogonal subspaces—while implementing a soft form of model predictive control.

Chapter 3: Continuous Behavior from Distinct Skills: Compositionality in Motor Cortex tests these theoretical predictions in motor cortex recordings from non-human primates performing a continuous force-tracking task. I find that motor cortex activity transitions from a condition-invariant preparatory regime to a dynamic execution regime, and that feedback perturbations engage the preparatory subspace even during movement. These findings provide empirical support for a re-planning interpretation of feedback-based correction and demonstrate that motor cortex flexibly deploys distinct neural subspaces to support planning and execution.

Chapter 4: Dual-Learning for Supervised Learning builds on this framework to address a central challenge in training large neural networks: how to balance fast, efficient learning with stability and long-term retention. I derive a theoretical bound on the maximum stable learning rate that explicitly captures the interaction between curvature and gradient noise. Motivated by this bound, I propose a dual-learning architecture in which a fast low-rank learner adapts quickly while a slow full-rank module consolidates long-term knowledge. This architecture enables efficient, robust learning, supports continual task acquisition, and aligns with biological motifs observed in thalamocortical loops.

Together, these studies advance a unified view of flexible motor behavior—one that integrates control, learning, and neurobiology—and lay the groundwork for scalable algorithms that mirror the brain’s capacity for adaptation and generalization.