Catastrophic Forgetting

The human brain achieves separation between rapid learning and long-term memory consolidation through the collaborative mechanism of the hippocampus and neocortex (complementary learning systems theory).

The catastrophic forgetting problem was first proposed by McCloskey and Cohen in 1989.

Catastrophic Forgetting is a classic pain point for neural networks in continual learning scenarios, where updating parameters on new tasks overwrites the weight distribution corresponding to old tasks, causing degradation of previously learned capabilities

The phenomenon of catastrophic forgetting was formally described by McCloskey and Cohen in 1989, referring to the tendency of neural networks to overwrite old knowledge when learning new tasks.

The three main strategies for mitigating catastrophic forgetting are regularization-based methods (such as EWC), replay-based methods (such as Experience Replay), and architecture-based methods (such as Progressive Neural Networks).

Catastrophic forgetting is a classic problem in the field of neural networks; as early as the 1990s, researchers discovered that the performance on old tasks drops sharply when neural networks are continuously trained on new tasks

An intermittently faulty memory module may alternately produce correct and incorrect computation results, causing sporadic deviations in gradient calculations during neural network training, affecting model convergence without triggering hardware alerts.

If only the data of newly added classes is used to fine-tune an expanded classification head, the model is highly prone to catastrophic forgetting, rapidly losing its ability to recognize the original classes.

Catastrophic forgetting refers to the phenomenon where a neural network rapidly overwrites old knowledge when learning new tasks; mainstream solutions in the research community include regularization methods (such as EWC), dynamic architecture methods, and replay methods.

Gradient-based methods often overwrite old knowledge when learning new tasks, causing catastrophic forgetting, whereas population-based evolutionary systems naturally maintain an archive of solution diversity.

In neural networks, parameters jointly encode knowledge of multiple tasks through gradient descent; when the model trains on a new task, gradient updates substantially change the weights critical to old tasks, while old task data no longer appears, causing the model to gradually forget the old tasks.

Representational reorganization refers to the phenomenon in which the structure of the network's internal representation space undergoes systematic adjustment when the task objective changes, typically manifesting as rapid flips in the ranking of dimension importance; it is distinct from catastrophic forgetting, which emphasizes the loss of performance on old tasks.

The developer has publicly released an experimental AI architecture called PNSN (Persistent Neural State Network), which attempts to solve catastrophic forgetting through parameter modularization and a trainable retrieval mechanism.

The framework avoids catastrophic forgetting when switching tasks by managing memory by name (the memory name is customizable).

Because of their few parameters and low computational overhead, lightweight models are more easily able to achieve rapid online fine-tuning on the robot body, but at the same time face risks of catastrophic forgetting and instability.

In this study, performing only mathematical reasoning distillation overwrites the financial knowledge the model previously acquired through vision-language alignment, although procedural mathematical skills are retained.

Traditional continual learning in the exemplar-free setting is prone to catastrophic forgetting, overwriting prior knowledge.

In test-time adaptation scenarios, agents need to continuously update their behavior based on new evidence during deployment while retaining old rules that remain valid; updating too aggressively or too conservatively both cause problems, and this trade-off is one of the important causes of the bottleneck revealed by AdaExploreBench.

SFT has a structural contradiction of catastrophic forgetting: the more deeply new data is learned, the more original capabilities are forgotten.

In the class-incremental graph learning scenario of graph neural networks, as new classes arrive one after another the model exhibits catastrophic forgetting, and predictions are often dominated by the most recently learned classes, producing recency bias.