Skip to yearly menu bar Skip to main content


Poster

Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training

Yanlai Yang · Matt Jones · Michael Mozer · Mengye Ren

West Ballroom A-D #5809
[ ]
Wed 11 Dec 11 a.m. PST — 2 p.m. PST

Abstract:

We explore the training dynamics of neural networks in a structured non-IID setting where documents are presented cyclically in a fixed, repeated sequence. Typically, networks suffer from catastrophic interference when training on a sequence of documents; however, we discover a curious and remarkable property of LLMs finetuned sequentially in this setting: they exhibit anticipatory behavior, recovering from the forgetting on documents before seeing them again. The behavior emerges and becomes more robust as the architecture scales up its number of parameters. Through comprehensive experiments and visualizations, we uncover new insights into training over-parameterized networks in structured environments.

Live content is unavailable. Log in and register to view live content