Continual Learning via Sparse Memory Finetuning

The paper addresses the challenge of continual learning in language models, focusing on mitigating catastrophic forgetting, where new data updates erase previously acquired knowledge. The authors propose a method called sparse memory finetuning, which involves updating only the memory slots highly activated by new knowledge, thereby reducing interference with existing capabilities. This approach is evaluated against full finetuning and parameter-efficient finetuning with LoRA on question answering tasks, demonstrating significantly less forgetting while maintaining new knowledge acquisition.

Sparse memory finetuning leverages memory layer models, which access a small set of parameters from a large memory pool during each forward pass. By updating only the top memory slots accessed more frequently in new data relative to pretraining data, the method minimizes interference with existing knowledge. The evaluation shows that while full finetuning and LoRA result in substantial forgetting, sparse memory finetuning achieves only an 11% drop in performance on NaturalQuestions, compared to 89% and 71% drops with full finetuning and LoRA, respectively.

The study highlights the potential of sparsity in memory layers for continual learning, suggesting that sparse updates can effectively balance learning new information and retaining existing knowledge. The authors also explore the trade-off between learning and forgetting, showing that sparse memory finetuning offers a Pareto improvement over other methods. The method's effectiveness is further supported by qualitative analyses of memory accesses, indicating that the TF-IDF ranking used for selecting trainable indices aligns with the semantic content of the knowledge being learned.

The paper acknowledges limitations, such as the need for further exploration of more sophisticated techniques for selecting sparse trainable parameters and scaling the approach to more complex tasks and larger models. Future work could also investigate input-dependent adaptation of the number of trainable indices and alternative ranking criteria to enhance the method's performance. Overall, the research demonstrates that sparse parameter updates in memory layers are a promising direction for achieving continual learning in language models without catastrophic forgetting.