Auto-Sizing Neural Networks: With Applications to n-gram Language Models

Abstract

Neural networks have been shown to improve performance across a range of natural-language tasks. However, designing and training them can be complicated. Frequently, researchers resort to repeated experimentation to pick optimal settings. In this paper, we address the issue of choosing the correct number of units in hidden layers. We introduce a method for automatically adjusting network size by pruning out hidden units through ${\ell }_{\infty ,1}$ and ${\ell }_{2,1}$ regularization. We apply this method to language modeling and demonstrate its ability to correctly choose the number of hidden units while maintaining perplexity. We also include these models in a machine translation decoder and show that these smaller neural models maintain the significant improvements of their unpruned versions.

EMNLP 2015

Cited by in corpus (9)

• Reducing Transformer Depth on Demand with Structured Dropout
• Sequence-Level Knowledge Distillation
• Model Selection in Bayesian Neural Networks via Horseshoe Priors
• Structured Variational Learning of Bayesian Neural Networks with Horseshoe Priors
• Computationally Efficient Neural Image Compression
• Compacting Neural Network Classifiers via Dropout Training
• Not All Attention Is All You Need
• Refining the Structure of Neural Networks Using Matrix Conditioning
• In-training Matrix Factorization for Parameter-frugal Neural Machine Translation