On Evaluating the Generalization of LSTM Models in Formal Languages

TL;DR

This paper empirically investigates how well LSTM models can learn and generalize formal languages, revealing significant differences based on training conditions and emphasizing careful evaluation of neural network capabilities.

Contribution

It provides an empirical analysis of LSTM generalization on formal languages, highlighting the impact of training regimes and model capacity on learning outcomes.

Findings

Performance varies significantly with training settings

Careful assessment is essential for claims about neural network capabilities

Different training data regimes influence generalization to unobserved samples

Abstract

Recurrent Neural Networks (RNNs) are theoretically Turing-complete and established themselves as a dominant model for language processing. Yet, there still remains an uncertainty regarding their language learning capabilities. In this paper, we empirically evaluate the inductive learning capabilities of Long Short-Term Memory networks, a popular extension of simple RNNs, to learn simple formal languages, in particular $a^nb^n$, $a^nb^nc^n$, and $a^nb^nc^nd^n$. We investigate the influence of various aspects of learning, such as training data regimes and model capacity, on the generalization to unobserved samples. We find striking differences in model performances under different training settings and highlight the need for careful analysis and assessment when making claims about the learning capabilities of neural network models.