Stable Anisotropic Regularization

arXiv:2305.19358v3 Announce Type: replace-cross
Abstract: Given the success of Large Language Models (LLMs), there has been considerable interest in studying the properties of model activations. The literature overwhelmingly agrees that LLM representations are dominated by a few "outlier dimensions" with exceedingly high variance and magnitude. Several studies in Natural Language Processing (NLP) have sought to mitigate the impact of such outlier dimensions and force LLMs to be isotropic (i.e., have uniform variance across all dimensions in embedding space). Isotropy is …

abstract arxiv cs.ai cs.cl cs.lg dimensions language language models language processing large language large language models literature llm llms natural natural language natural language processing nlp outlier processing regularization studies studying success type variance