Paper: https://arxiv.org/abs/2311.10642
Code: https://github.com/vulus98/Rethinking-attention
Abstract:
This work presents an analysis of the effectiveness of using standard shallow feed-forward networks to mimic the behavior of the attention mechanism in the original Transformer model, a state-of-the-art architecture for sequence-to-sequence tasks. We substitute key elements of the attention mechanism in the Transformer with simple feed-forward networks, trained using the original components via knowledge distillation. Our experiments, conducted on the IWSLT2017 dataset, reveal the capacity of these “attentionless Transformers” to rival the performance of the original architecture. Through rigorous ablation studies, and experimenting with various replacement network types and sizes, we offer insights that support the viability of our approach. This not only sheds light on the adaptability of shallow feed-forward networks in emulating attention mechanisms but also underscores their potential to streamline complex architectures for sequence-to-sequence tasks.
The reasons this isn’t done:
- Fixed max sequence length (shorter sequences aren’t less computation)
- Very short max sequence length (50 tokens in this paper!)
- Very inefficient training (for a target sequence with N tokens, this requires N forward passes for the decoder, as opposed to 1 with attention, because there’s no obvious way to parallelize the causal self-attention with a FF
because there’s no obvious way to parallelize the causal self-attention with a FF
You can just use triangular matrices, autoregressive language modelling can be done even with linear only layers. See page 12 of https://arxiv.org/abs/2309.06979