Sparsity Emerges Naturally in Neural Language Models

Abstract

Concerns about interpretability, computational resources, and principled inductive priors have motivated efforts to engineer sparse neural models for NLP tasks. If sparsity is important for NLP, might well-trained neural models naturally become roughly sparse? Using the Taxi-Euclidean norm to measure sparsity, we find that frequent input words are associated with concentrated or sparse activations, while frequent target words are associated with dispersed activations but concentrated gradients. We find that gradients associated with function words are more concentrated than the gradients of content words, even controlling for word frequency.

arXiv commentPublished in the ICML 2019 Workshop on Identifying and Understanding Deep Learning Phenomena: https://openreview.net/forum?id=H1ets1h56E

References in corpus (11)

• Visualizing and Understanding Recurrent Networks
• Opening the Black Box of Deep Neural Networks via Information
• Understanding Neural Networks through Representation Erasure
• Learning to Generate Reviews and Discovering Sentiment
• Visualizing and Understanding Neural Models in NLP
• Exploring Sparsity in Recurrent Neural Networks
• Euclid in a Taxicab: Sparse Blind Deconvolution with Smoothed l1/l2 Regularization
• Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context
• Representation of linguistic form and function in recurrent neural networks
• Compressing Neural Language Models by Sparse Word Representations
• A Deep Generative Model of Vowel Formant Typology