Catastrophic Forgetting: How Fine-Tuning AI Models Can Erase General Knowledge

Fine-tuning AI models on specific tasks often causes 'catastrophic forgetting,' where the model gains specialised skills but loses broader general knowledge. This happens because adjusting a model's weights for a new task tends to overwrite previously learned patterns, given the model's limited capacity. Techniques such as regularisation, rehearsal with old data, multi-task learning, and parameter-efficient methods like LoRA can reduce but not fully eliminate this trade-off. For many real-world applications, however, specialisation may outweigh the cost of reduced generality. Researchers expect advances in continuous and meta-learning to gradually address forgetting over the next decade.