Component-Aware Pruning Framework for Neural Network Controllers via Gradient-Based Importance Estimation

TL;DR

This paper presents a gradient-based, component-aware pruning framework for neural network controllers that improves model compression by capturing functional importance and structural dependencies during training.

Contribution

It introduces a novel importance estimation method using gradient information, enabling more effective and informed pruning of neural network components.

Findings

Reveals critical structural dependencies in neural controllers.

Captures dynamic importance shifts during training.

Supports more effective model compression decisions.

Abstract

The transition from monolithic to multi-component neural architectures in advanced neural network controllers poses substantial challenges due to the high computational complexity of the latter. Conventional model compression techniques for complexity reduction, such as structured pruning based on norm-based metrics to estimate the relative importance of distinct parameter groups, often fail to capture functional significance. This paper introduces a component-aware pruning framework that utilizes gradient information to compute three distinct importance metrics during training: Gradient Accumulation, Fisher Information, and Bayesian Uncertainty. Experimental results with an autoencoder and a TD-MPC agent demonstrate that the proposed framework reveals critical structural dependencies and dynamic shifts in importance that static heuristics often miss, supporting more informed compression decisions.