MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head
This addresses the computational bottleneck of Transformers for large-scale applications by providing an efficient linear attention method with significant performance gains.
The paper tackled the performance degradation of linear attention in Transformers by proposing Multi-Head Linear Attention (MHLA) to restore expressivity, achieving improvements such as 3.6% on ImageNet classification and 41% on video generation.
While the Transformer architecture dominates many fields, its quadratic self-attention complexity hinders its use in large-scale applications. Linear attention offers an efficient alternative, but its direct application often degrades performance, with existing fixes typically re-introducing computational overhead through extra modules (e.g., depthwise separable convolution) that defeat the original purpose. In this work, we identify a key failure mode in these methods: global context collapse, where the model loses representational diversity. To address this, we propose Multi-Head Linear Attention (MHLA), which preserves this diversity by computing attention within divided heads along the token dimension. We prove that MHLA maintains linear complexity while recovering much of the expressive power of softmax attention, and verify its effectiveness across multiple domains, achieving a 3.6\% improvement on ImageNet classification, a 6.3\% gain on NLP, a 12.6\% improvement on image generation, and a 41\% enhancement on video generation under the same time complexity.