SAMSA: Efficient Transformer for Many Data Modalities

The versatility of self-attention mechanism earned transformers great success in almost all data modalities, with limitations on the quadratic complexity and difficulty of training. Efficient transformers, on the other hand, often rely on clever data-modality-dependent construction to get over the quadratic complexity of transformers. This greatly hinders their applications on different data modalities, which is one of the pillars of contemporary foundational modeling. In this paper, we lay the groundwork for efficient foundational modeling by proposing SAMSA - SAMpling-Self-Attention, a context-aware linear complexity self-attention mechanism that works well on multiple data modalities. Our mechanism is based on a differentiable sampling without replacement method we discovered. This enables the self-attention module to attend to the most important token set, where the importance is defined by data. Moreover, as differentiability is not needed in inference, the sparse formulation of our method costs little time overhead, further lowering computational costs. In short, SAMSA achieved competitive or even SOTA results on many benchmarks, while being faster in inference, compared to other very specialized models. Against full self-attention, real inference time significantly decreases while performance ranges from negligible degradation to outperformance. We release our source code in the repository: https://github.com/HySonLab/SAMSA