Date of Award
Fall 1-1-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Biomedical Engineering (ENAS)
First Advisor
Saxena, Shreya
Abstract
Understanding how populations of neurons generate behavior remains a fundamental question in neuroscience. While single-neuron analyses have provided valuable insights, behavior emerges from the coordinated dynamics of large neural ensembles that evolve over time and across states. A major difficulty is that neural population activity is non-stationary: it exhibits temporal distributional shifts, transitions between distinct dynamical regimes, and modulation by external or inter-regional inputs. Advances in artificial intelligence offer new opportunities to address this challenge. Yet, conventional models, such as standard recurrent neural networks, often fail to capture such complexity, limiting both predictive performance and mechanistic interpretability. This motivates the development of new modeling frameworks. This dissertation introduces three frameworks based on recurrent neural networks to address these challenges. First, Time-varying Recurrent Neural Networks (TV-RNNs) adapt their recurrent weights over time, enabling robust early classification of behavior from neural activity despite temporal distributional shifts. Second, Switching Recurrent Neural Networks (SRNNs) model neural activity as transitions among low-dimensional nonlinear dynamical regimes governed by Markov switching, allowing the direct recovery of behaviorally relevant neural states and their temporal structure. Finally, Input-driven Switching Recurrent Neural Networks (iSRNNs) extend this framework by explicitly disentangling intrinsic dynamics from input-driven modulations, uncovering interpretable drivers of neural state transitions and inter-regional signals that shape these dynamics. Together, these contributions establish frameworks for linking behavior with neural dynamics. By combining temporally adaptive, switching, and input-driven architectures, this work improves both decoding performance and mechanistic interpretability, offering new perspectives on how stable intrinsic motifs and flexible input-driven perturbations shape neural computation. Beyond providing methodological advances, the findings suggest future works for studying cross-region communication, hierarchical organization, and causal mechanisms underlying neural state transitions.
Recommended Citation
Zhang, Yongxu, "Bridging Behavior and Neural Activity: Neural Dynamics Leading to Self-Initiated and Task-Driven Movements" (2025). Yale Graduate School of Arts and Sciences Dissertations. 1878.
https://elischolar.library.yale.edu/gsas_dissertations/1878