Training Thinner and Deeper Neural Networks: Jumpstart Regularization

TL;DR

This paper introduces jumpstart regularization, a novel technique to train thinner and deeper neural networks more efficiently by preventing neuron death and linearization, thus improving parameter efficiency and depth without excessive computational costs.

Contribution

The paper proposes jumpstart regularization, a new method that enables training of deeper, thinner neural networks with better parameter efficiency compared to traditional approaches.

Findings

Neural networks trained with jumpstart regularization are thinner and deeper.

Jumpstart regularization prevents neurons from dying or becoming linear.

The method improves parameter efficiency over conventional training.

Abstract

Neural networks are more expressive when they have multiple layers. In turn, conventional training methods are only successful if the depth does not lead to numerical issues such as exploding or vanishing gradients, which occur less frequently when the layers are sufficiently wide. However, increasing width to attain greater depth entails the use of heavier computational resources and leads to overparameterized models. These subsequent issues have been partially addressed by model compression methods such as quantization and pruning, some of which relying on normalization-based regularization of the loss function to make the effect of most parameters negligible. In this work, we propose instead to use regularization for preventing neurons from dying or becoming linear, a technique which we denote as jumpstart regularization. In comparison to conventional training, we obtain neural networks that are thinner, deeper, and - most importantly - more parameter-efficient.