Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training

arXiv:2403.09613v1 Announce Type: new
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 fine-tuned sequentially in this setting: they exhibit anticipatory behavior, recovering from the forgetting on documents before encountering them again. The behavior emerges and becomes more robust as the architecture …

abstract arxiv cs.cl cs.lg documents dynamics explore however interference knowledge networks neural networks property recovery training type via