Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay
Martin Marek, Dongkyu Cho, Shikai Qiu, Rumi Chunara, Pavel Izmailov, Andrew Gordon Wilson
Key claim
Self-generated samples nearly eliminate forgetting in language models.
This paper addresses the issue of forgetting in language models when trained on new tasks. It shows that self-generated samples can effectively serve as replay data, significantly reducing forgetting. The key result is that this method allows for high-learning-rate finetuning without the typical tradeoff of forgetting.
The paper introduces a novel approach to mitigate forgetting in language models using self-generated samples.
The findings are supported by experimental results demonstrating the effectiveness of the proposed method.
Deep reliability assessment
The methodology supports the claim that self-generated replay data can effectively mitigate forgetting in language models, but it may overclaim the extent to which this approach can be generalized across all model architectures and tasks.
Reproducibility
Yes, the code is available at https://github.com/martin-marek/forgetting.
Discussion questions
- What assumptions are made about the model's capacity that could limit the applicability of these findings?
- How can builders implement self-generated replay in their own models without extensive computational resources?
- What specific conditions or experiments would demonstrate that self-generated replay does not prevent forgetting?
Key figure
Figure 1 illustrates the relationship between model capacity and forgetting, showing that smaller models trained close to saturation struggle to retain knowledge from prior tasks during finetuning.