Generating Long Sequences with Sparse Transformers
This addresses the scalability problem for long-sequence modeling in AI, enabling more efficient processing across domains like images, audio, and text, though it is incremental as it builds on existing Transformer architectures.
The paper tackles the quadratic time and memory complexity of Transformers by introducing sparse factorizations of the attention matrix, reducing it to O(n√n), and demonstrates the ability to model sequences tens of thousands of timesteps long, setting new state-of-the-art results on density modeling benchmarks like Enwik8, CIFAR-10, and ImageNet-64.
Transformers are powerful sequence models, but require time and memory that grows quadratically with the sequence length. In this paper we introduce sparse factorizations of the attention matrix which reduce this to $O(n \sqrt{n})$. We also introduce a) a variation on architecture and initialization to train deeper networks, b) the recomputation of attention matrices to save memory, and c) fast attention kernels for training. We call networks with these changes Sparse Transformers, and show they can model sequences tens of thousands of timesteps long using hundreds of layers. We use the same architecture to model images, audio, and text from raw bytes, setting a new state of the art for density modeling of Enwik8, CIFAR-10, and ImageNet-64. We generate unconditional samples that demonstrate global coherence and great diversity, and show it is possible in principle to use self-attention to model sequences of length one million or more.