RNN Architecture Learning with Sparse Regularization

TL;DR

This paper introduces a sparse regularization method for rational RNNs that reduces model size significantly while maintaining performance, enhancing interpretability and efficiency in NLP tasks.

Contribution

The authors develop a group lasso-based structure learning approach for rational RNNs, enabling substantial parameter reduction without performance loss.

Findings

Models with over 90% weight pruning perform comparably to dense models.

Sparse rational RNNs are more interpretable and easier to visualize.

The method is effective across multiple sentiment analysis datasets with different embeddings.

Abstract

Neural models for NLP typically use large numbers of parameters to reach state-of-the-art performance, which can lead to excessive memory usage and increased runtime. We present a structure learning method for learning sparse, parameter-efficient NLP models. Our method applies group lasso to rational RNNs (Peng et al., 2018), a family of models that is closely connected to weighted finite-state automata (WFSAs). We take advantage of rational RNNs' natural grouping of the weights, so the group lasso penalty directly removes WFSA states, substantially reducing the number of parameters in the model. Our experiments on a number of sentiment analysis datasets, using both GloVe and BERT embeddings, show that our approach learns neural structures which have fewer parameters without sacrificing performance relative to parameter-rich baselines. Our method also highlights the interpretable properties of rational RNNs. We show that sparsifying such models makes them easier to visualize, and we present models that rely exclusively on as few as three WFSAs after pruning more than 90% of the weights. We publicly release our code.