Sparsifying Transformer Models with Trainable Representation Pooling
This addresses efficiency bottlenecks in large-scale Transformer models for NLP tasks like summarization, offering significant speed-ups with minimal quality loss.
The paper tackles the quadratic complexity of Transformer attention by introducing a trainable method to sparsify attention, focusing on task-specific tokens, resulting in up to 13x computational efficiency gains while maintaining state-of-the-art quality on long document summarization.
We propose a novel method to sparsify attention in the Transformer model by learning to select the most-informative token representations during the training process, thus focusing on the task-specific parts of an input. A reduction of quadratic time and memory complexity to sublinear was achieved due to a robust trainable top-$k$ operator. Our experiments on a challenging long document summarization task show that even our simple baseline performs comparably to the current SOTA, and with trainable pooling, we can retain its top quality, while being $1.8\times$ faster during training, $4.5\times$ faster during inference, and up to $13\times$ more computationally efficient in the decoder.