Long Range Language Modeling via Gated State Spaces
This work addresses long-range dependency modeling for language and code tasks, offering a faster and competitive alternative to Transformers, though it appears incremental as it builds on existing state space models and gating techniques.
The authors tackled long-range autoregressive sequence modeling across English books, GitHub code, and ArXiv math articles by proposing a Gated State Space (GSS) layer, which trains faster than DSS on TPUs, is competitive with Transformers, and generalizes zero-shot to longer inputs.
State space models have shown to be effective at modeling long range dependencies, specially on sequence classification tasks. In this work we focus on autoregressive sequence modeling over English books, Github source code and ArXiv mathematics articles. Based on recent developments around the effectiveness of gated activation functions, we propose a new layer named Gated State Space (GSS) and show that it trains significantly faster than the diagonal version of S4 (i.e. DSS) on TPUs, is fairly competitive with several well-tuned Transformer-based baselines and exhibits zero-shot generalization to longer inputs while being straightforward to implement. Finally, we show that leveraging self-attention to model local dependencies improves the performance of GSS even further.