Functional Interpolation for Relative Positions Improves Long Context Transformers
This addresses the challenge of extending context length in Transformers for applications like long-form text processing, though it appears incremental as it builds on existing relative position encodings.
The paper tackles the problem of Transformer performance decay on inputs longer than training sequences by proposing FIRE, a functional relative position encoding with progressive interpolation, which improves generalization to longer contexts in zero-shot language modeling and long text benchmarks.
Preventing the performance decay of Transformers on inputs longer than those used for training has been an important challenge in extending the context length of these models. Though the Transformer architecture has fundamentally no limits on the input sequence lengths it can process, the choice of position encoding used during training can limit the performance of these models on longer inputs. We propose a novel functional relative position encoding with progressive interpolation, FIRE, to improve Transformer generalization to longer contexts. We theoretically prove that this can represent some of the popular relative position encodings, such as T5's RPE, Alibi, and Kerple. We next empirically show that FIRE models have better generalization to longer contexts on both zero-shot language modeling and long text benchmarks.