Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings
This addresses a fundamental limitation in Transformer architectures for researchers and practitioners, offering a more robust positional encoding method that enhances model performance and generalization.
The paper tackles the entanglement of content and position in RoPE positional embeddings in Transformers, proposing PoPE to decouple them, which improves performance on diagnostic tasks and autoregressive modeling across music, genomic, and language domains, with gains in perplexity and zero-shot length extrapolation.
The attention mechanism in a Transformer architecture matches key to query based on both content -- the what -- and position in a sequence -- the where. We present an analysis indicating that what and where are entangled in the popular RoPE rotary position embedding. This entanglement can impair performance particularly when decisions require independent matches on these two factors. We propose an improvement to RoPE, which we call Polar Coordinate Position Embeddings or PoPE, that eliminates the what-where confound. PoPE is far superior on a diagnostic task requiring indexing solely by position or by content. On autoregressive sequence modeling in music, genomic, and natural language domains, Transformers using PoPE as the positional encoding scheme outperform baselines using RoPE with respect to evaluation loss (perplexity) and downstream task performance. On language modeling, these gains persist across model scale, from 124M to 774M parameters. Crucially, PoPE shows strong zero-shot length extrapolation capabilities, whereas RoPE's performance degrades significantly on longer sequences at test time without fine tuning or the use of position-interpolation methods.