KERPLE: Kernelized Relative Positional Embedding for Length Extrapolation

Ta-Chung Chi, Ting-Han Fan, Peter Ramadge, Alexander Rudnicky

Keywords: Transformer Language Modeling, Length Extrapolation, Kernel Method

Abstract: Relative positional embeddings (RPE) have received considerable attention since RPEs effectively model the relative distance among tokens and enable length extrapolation. We propose KERPLE, a framework that generalizes relative position embedding for extrapolation by kernelizing positional differences. We achieve this goal using conditionally positive definite (CPD) kernels, a class of functions known for generalizing distance metrics. To maintain the inner product interpretation of self-attention, we show that a CPD kernel can be transformed into a PD kernel by adding a constant offset. This offset is implicitly absorbed in the Softmax normalization during self-attention. The diversity of CPD kernels allows us to derive various RPEs that enable length extrapolation in a principled way. Experiments demonstrate that the logarithmic variant achieves excellent extrapolation performance on three large language modeling datasets. Our implementation and pretrained checkpoints are released at~\url{https://github.com/chijames/KERPLE.git}.

TL;DR: We showed that conditionally positive definite (CPD) kernels allow us to derive various relative positional embeddings (RPE) with superior performance on the task of transformer language modeling length extrapolation.