On Difficulties of Attention Factorization through Shared Memory

arXiv:2404.00798v1 Announce Type: new
Abstract: 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 …

abstract arxiv attention audio challenges complexity computer computer vision cs.lg deep learning factorization fields however inputs language language processing lies memory natural natural language natural language processing processing relationships researchers through transformers type vision