Adaptive Neural Networks Using Residual Fitting

TL;DR

This paper introduces a neural network growth method that adaptively increases network capacity based on residual error, achieving comparable or better performance with less computational effort than traditional architecture search or pruning.

Contribution

The paper proposes a novel residual fitting-based network-growth approach that adaptively determines network size during training, reducing the need for extensive architecture search.

Findings

Growing networks outperform small fixed networks in various tasks.

The method achieves similar or better performance compared to larger static networks.

Network growth based on residual error is effective across classification, imitation learning, and reinforcement learning.

Abstract

Current methods for estimating the required neural-network size for a given problem class have focused on methods that can be computationally intensive, such as neural-architecture search and pruning. In contrast, methods that add capacity to neural networks as needed may provide similar results to architecture search and pruning, but do not require as much computation to find an appropriate network size. Here, we present a network-growth method that searches for explainable error in the network's residuals and grows the network if sufficient error is detected. We demonstrate this method using examples from classification, imitation learning, and reinforcement learning. Within these tasks, the growing network can often achieve better performance than small networks that do not grow, and similar performance to networks that begin much larger.