← Search

Pulkit Gopalani

2 accepted papers

2024

Abrupt Learning in Transformers: A Case Study on Matrix Completion

NeurIPS 2024poster

Recent analysis on the training dynamics of Transformers has unveiled an interesting characteristic: the training loss plateaus for a significant number of training steps, and then suddenly (and sharply) drops to near--optimal values. To understand this phenomenon in depth, we formulate the low-rank…

Cited by 1SourcePDFScholar