On Difficulties of Attention Factorization through Shared Memory
This work addresses efficiency bottlenecks in transformers for large-scale AI applications, offering an incremental improvement over existing memory-augmented methods.
The paper challenges the conventional approach of directly interfacing attention with external memory in models like Luna, showing that filtering the input before memory communication significantly improves performance.
Transformers have revolutionized deep learning in numerous fields, including natural language processing, computer vision, and audio processing. Their strength lies in their attention mechanism, which allows for the discovering of complex input relationships. However, this mechanism's quadratic time and memory complexity pose challenges for larger inputs. Researchers are now investigating models like Linear Unified Nested Attention (Luna) or Memory Augmented Transformer, which leverage external learnable memory to either reduce the attention computation complexity down to linear, or to propagate information between chunks in chunk-wise processing. Our findings challenge the conventional thinking on these models, revealing that interfacing with the memory directly through an attention operation is suboptimal, and that the performance may be considerably improved by filtering the input signal before communicating with memory.