Towards the Renormalization of Transformer Language Models

Transformer language models (LMs), the cornerstone of the popular large language models (LLMs), have revolutionised the fields of artificial intelligence (AI) and natural language processing (NLP). However, we still understand relatively little about their computational mechanisms and the reasons for their effectiveness. The internal workings of these models remain enigmatic "black boxes." In this thesis, I present a framework for interpreting LMs, which can be aptly described using the term "renormalization." Inherent model components — such as layers, attention heads, MLPs, or neurons — should serve as the base units of abstraction for the renormalization process. This approach is preferable to using familiar linguistic or cognitive structures, as there is insufficient evidence to suggest that the internal mechanisms of LMs operate in a manner analogous to human linguistic or cognitive frameworks. Furthermore, our analysis of the Knowledge Neuron (KN) thesis suggests that neurons exhibit a remarkable degree of idiosyncrasy in the types of function and information they associate with. This finding underscores the necessity of conducting the coarse-graining process at the granularity of individual neurons. However, our examination also reveals critical limitations of the KN thesis, suggesting that simply decomposing the model without accounting for interactions between these units provides an incomplete explanation. These interactions can be captured through circuit discovery methods. To address this, we propose a novel approach that enables joint pruning of connection edges and weight parameters, facilitating efficient neuron-level granularity. As a result, our method surpasses previous approaches, achieving state-of-the-art results in LM interpretation. Finally, our findings reveal two key insights for applications: global representational techniques that involve modifications to the entire model should be prioritised, while conventional, pre-theoretical speculation from cognitive science and linguistics has to date not been a fruitful enterprise. To demonstrate this, we present a case study on the task of temporal relation extraction, where, by leveraging these insights, we propose two effective improvements that achieve state-of-the-art performance.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.