{"title":"分层注意力快捷方式自适应大型语言模型","authors":"Prateek Verma, Mert Pilanci","doi":"arxiv-2409.10870","DOIUrl":null,"url":null,"abstract":"Transformer architectures are the backbone of the modern AI revolution.\nHowever, they are based on simply stacking the same blocks in dozens of layers\nand processing information sequentially from one block to another. In this\npaper, we propose to challenge this and introduce adaptive computations for\nLLM-like setups, which allow the final layer to attend to all of the\nintermediate layers as it deems fit through the attention mechanism, thereby\nintroducing computational \\textbf{attention shortcuts}. These shortcuts can\nthus make the architecture depth and context adaptive. We showcase four\ndifferent datasets, namely acoustic tokens, natural language, and symbolic\nmusic, and we achieve superior performance for GPT-like architecture. We give\nevidence via attention maps that the models learn complex dependencies across\nlayers that are adaptive in context and depth depending on the input tokens.","PeriodicalId":501284,"journal":{"name":"arXiv - EE - Audio and Speech Processing","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2024-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Adaptive Large Language Models By Layerwise Attention Shortcuts\",\"authors\":\"Prateek Verma, Mert Pilanci\",\"doi\":\"arxiv-2409.10870\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Transformer architectures are the backbone of the modern AI revolution.\\nHowever, they are based on simply stacking the same blocks in dozens of layers\\nand processing information sequentially from one block to another. In this\\npaper, we propose to challenge this and introduce adaptive computations for\\nLLM-like setups, which allow the final layer to attend to all of the\\nintermediate layers as it deems fit through the attention mechanism, thereby\\nintroducing computational \\\\textbf{attention shortcuts}. These shortcuts can\\nthus make the architecture depth and context adaptive. We showcase four\\ndifferent datasets, namely acoustic tokens, natural language, and symbolic\\nmusic, and we achieve superior performance for GPT-like architecture. We give\\nevidence via attention maps that the models learn complex dependencies across\\nlayers that are adaptive in context and depth depending on the input tokens.\",\"PeriodicalId\":501284,\"journal\":{\"name\":\"arXiv - EE - Audio and Speech Processing\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-09-17\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"arXiv - EE - Audio and Speech Processing\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/arxiv-2409.10870\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - EE - Audio and Speech Processing","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.10870","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Adaptive Large Language Models By Layerwise Attention Shortcuts
Transformer architectures are the backbone of the modern AI revolution.
However, they are based on simply stacking the same blocks in dozens of layers
and processing information sequentially from one block to another. In this
paper, we propose to challenge this and introduce adaptive computations for
LLM-like setups, which allow the final layer to attend to all of the
intermediate layers as it deems fit through the attention mechanism, thereby
introducing computational \textbf{attention shortcuts}. These shortcuts can
thus make the architecture depth and context adaptive. We showcase four
different datasets, namely acoustic tokens, natural language, and symbolic
music, and we achieve superior performance for GPT-like architecture. We give
evidence via attention maps that the models learn complex dependencies across
layers that are adaptive in context and depth depending on the input tokens.