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

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.

To achieve continuous growth, the problem of 'catastrophic forgetting' must be solved - training on new data overwrites old knowledge.

ZY-S1's MATH-500 dropping from 4 to 0 is a typical manifestation of catastrophic forgetting caused by a distribution shift in the fine-tuning data.

In continual learning scenarios, if one cannot distinguish which information has been consolidated into the parameters and which is merely state drift, it is difficult to determine at which layer catastrophic forgetting actually occurs.

In recurrent networks, the catastrophic forgetting problem is amplified: because hidden states are recursively coupled over time, a small change in the weights may be amplified exponentially through multi-step propagation.

The post-training stage carries the risk of catastrophic forgetting, and post-training of web-native models usually adopts a mixed-data strategy, training on a blend of general instruction data and web interaction data.

When adapting the general-domain trained Whisper-small model to the Earnings22 benchmark dataset, all evaluation methods showed improvement rather than degradation on the in-domain held-out set of financial speech during the adaptation process, and the expected catastrophic forgetting did not occur.

The root cause of catastrophic forgetting is that the optimal weight regions for the two tasks do not overlap, and the optimization directions are in fundamental conflict.

Under naive training, the network is first trained to 100% accuracy on Task A, then trained only on Task B data, after which Task A accuracy drops from 100% to 64% while Task B reaches 100%.

Fine-tuning adult ASR models on children's speech data causes catastrophic forgetting, degrading the model's recognition capability on adult speech.

Interleaved training can prevent catastrophic forgetting, stopping the model from forgetting previously learned data when learning new types of data.

Supervised fine-tuning (SFT) tends to cause more severe catastrophic forgetting than reinforcement fine-tuning (RFT).