))
LLM DIFF Transformer with SoftMax Subtraction
Manage episode 444738222 series 3605861
Innehåll tillhandahållet av Brian Carter. Allt poddinnehåll inklusive avsnitt, grafik och podcastbeskrivningar laddas upp och tillhandahålls direkt av Brian Carter eller deras podcastplattformspartner. Om du tror att någon använder ditt upphovsrättsskyddade verk utan din tillåtelse kan du följa processen som beskrivs här https://sv.player.fm/legal.
This paper presents a new architecture for large language models called DIFF Transformer. The paper argues that conventional Transformers over-allocate attention to irrelevant parts of the input, drowning out the signal needed for accurate output. DIFF Transformer tackles this issue by using a differential attention mechanism that subtracts two softmax attention maps, effectively canceling out noise and amplifying attention to relevant content. The paper presents extensive experiments demonstrating that DIFF Transformer outperforms conventional Transformers in various tasks, including language modeling, key information retrieval, hallucination mitigation, and in-context learning. This results in a more efficient model that requires fewer parameters and training data to achieve the same performance as a Transformer.
Read more: https://arxiv.org/pdf/2410.05258
71 episoder
Dela
