Ganesh Sundaram

• Neural networks are nowadays fundamental for modeling complex dynamical systems and are widely used for system identification and feedback control. However, as model capacity increases, deploying these networks on embedded edge devices is challenging due to limited computational and memory resources. Additionally, control applications impose stringent real-time constraints and require that closed-loop stability be preserved after model compression. Model compression is therefore essential. However, most existing methods are designed for general-purpose, monolithic networks and do not translate well to neural-network controllers. Modern controllers are often implemented as Multi-Component Neural Architectures (MCNAs), in which specialized subnetworks interact through coupling interfaces and often process latent representations derived from high-dimensional observations. Uniform, structure-agnostic sparsification may overlook these dependencies, thereby degrading closed-loop performance. This dissertation introduces a component-aware structured pruning framework tailored for multi-component neural architectures. The proposed approach decomposes an MCNA into module-specific dependency graphs, distinguishing component-internal parameter groups from inter-component coupling groups to facilitate targeted pruning. Group-wise pruning rates are determined by casting sparsity allocation as a constrained search problem, solved via grid search and continuous optimization to achieve compression goals while minimizing task performance degradation. Furthermore, both post-training and training-aware group-importance estimation methods are developed. The training-aware method computes online importance signals using gradient accumulation, diagonal Fisher information, and an empirical Bayesian mechanism to capture both immediate sensitivity and long-term learning dynamics. The framework is evaluated on representative architectures, including an MNIST autoencoder and a TD-MPC agent. Results demonstrate that integrating application-specific metrics into the compression objective enables more precise control over the pruning distribution and achieves greater reductions in model size and inference latency than baseline pruning methods while preserving closed-loop performance and stability requirements in the evaluated scenarios.