Date of Award

Fall 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering (ENAS)

First Advisor

Saxena, Shreya

Abstract

Understanding the relationship between neural activity and behavior is fundamental to uncovering brain function. With recent advancements in hardware and computation, we can now collect and analyze high-resolution behavioral and neural data in complex, naturalistic settings. However, challenges remain in effectively modeling and disentangling high-dimensional behaviors, particularly in social and multi-subject contexts, where subtle variations may reflect individual differences or underlying neural states. This thesis presents a series of structured, interpretable deep learning frameworks that address this challenge by modeling behavior and its neural correlates across multiple levels. First, I introduce the Constrained Subspace Variational Autoencoder (CS-VAE), which uses semi-supervised learning and Cauchy-Schwarz divergence to disentangle shared versus individual features in multi-subject and social behavioral datasets. This model enables the identification of interpretable behavioral motifs, and generalizes across subjects in naturalistic environments. Next, I extend this approach to uncover hierarchical social behavior motifs, enabling an unbiased and fine-grained analysis of how animals interact. Using a constrained latent space, the model captures both across-behavior class and within-class variability, and reveals how genotype (e.g., 16p11.2^dp/+ vs. wildtype) modulates behavior even within the same stereotyped motif. Finally, I present Shared-AE, a multimodal autoencoder framework that aligns latent subspaces between behavioral and neural recordings while preserving modality-specific features. This approach reveals how neural activity encodes complex behaviors and identifies shared motifs across neural and behavioral modalities. Across all three works, I demonstrate that these models not only improve interpretability and generalization in behavioral analysis, but also enable a deeper understanding of the neural basis of individual and social behavior. Future directions include incorporating temporal attention or transformer-based models to better capture long-range behavioral dependencies, integrating physiological signals (e.g., heart rate, respiration) for richer multimodal analyses, and extending the Shared-AE framework to heterogeneous neural datasets with varying temporal resolutions. These extensions promise to deepen our understanding of behavior across timescales and modalities, supporting comprehensive models of brain-behavior relationships.

Download

Share

COinS