Catastrophic Forgetting
Catastrophic forgetting describes the phenomenon where artificial neural networks, upon learning new tasks, lose previously acquired knowledge. Current research focuses on mitigating this issue through various strategies, including parameter-efficient fine-tuning methods (like LoRA), generative model-based data replay, and novel optimization algorithms that constrain gradient updates or leverage hierarchical task structures. Addressing catastrophic forgetting is crucial for developing robust and adaptable AI systems capable of continuous learning in real-world applications, particularly in domains like medical imaging, robotics, and natural language processing where data streams are constantly evolving.
Papers
On Reinforcement Learning and Distribution Matching for Fine-Tuning Language Models with no Catastrophic Forgetting
Tomasz Korbak, Hady Elsahar, Germán Kruszewski, Marc Dymetman
Analysis of Catastrophic Forgetting for Random Orthogonal Transformation Tasks in the Overparameterized Regime
Daniel Goldfarb, Paul Hand
Transfer without Forgetting
Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, Simone Calderara