Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to abruptly and drastically forget previously learned information upon learning new information. This phenomenon is a central challenge in Machine learning and Deep learning, particularly when models are trained sequentially on multiple tasks. It is a radical manifestation of the stability-plasticity dilemma, which refers to the difficulty of making a network that is sensitive to new information without being disrupted by it.
Neural networks are a key component of the connectionist approach to cognitive science. The issue of catastrophic interference in modeling human memory was first brought to scientific attention by research from McCloskey and Cohen (1989) and Ratcliff (1990). These studies highlighted a fundamental difference between artificial networks and human memory: while humans typically do not exhibit such drastic forgetting, standard backpropagation networks often do.
The Stability-Plasticity Dilemma
The stability-plasticity dilemma describes the trade-off between a system's ability to integrate new information (plasticity) and its ability to preserve existing knowledge (stability). Lookup tables and connectionist networks lie on opposite ends of this spectrum. A lookup table remains completely stable when new information is added but cannot generalize to unseen inputs. In contrast, connectionist networks, such as standard backpropagation networks, can generalize to novel inputs but are highly sensitive to new information, often leading to catastrophic interference.
This sensitivity is problematic when modeling human memory, because humans generally retain previously learned information while acquiring new knowledge. The contrast suggests that biological memory systems employ mechanisms that artificial networks lack, such as complementary learning systems or synaptic consolidation.
Sequential Learning Experiments
McCloskey and Cohen (1989) demonstrated catastrophic interference through two experiments with backpropagation networks. In the first experiment, they trained a network on single-digit addition facts involving the number one (e.g., 1+1 through 9+1 and 1+2 through 1+9) until it responded correctly to all. They then trained the same network on addition facts involving the number two (e.g., 2+1 through 2+9 and 1+2 through 9+2). After just one training trial on the twos facts, the network's performance on the ones facts dropped dramatically, often producing outputs resembling incorrect answers. Even overlapping problems like 2+1 and 1+2 showed significant disruption.
In a second experiment, McCloskey and Cohen replicated a classic study on retroactive interference in humans by Barnes and Underwood (1959). They trained a network on paired associates (A-B and A-C lists) with context patterns to distinguish the lists. When trained concurrently on both lists, the network learned all associations correctly. However, when trained sequentially (A-B first, then A-C), the amount of A-C training that produced 50% correct responses in humans led to nearly 0% correct responses for the A-B associations in the network. The network's outputs for B prompts often resembled C responses, indicating that new learning had overwritten old knowledge.
McCloskey and Cohen attempted to reduce interference through various manipulations, including changing the number of hidden units, adjusting the learning rate, overtraining on the A-B list, freezing certain connection weights, and altering target values. None of these satisfactorily mitigated the catastrophic forgetting.
Recognition Memory Findings
Ratcliff (1990) extended these findings using backpropagation models applied to standard recognition memory procedures, where items were learned sequentially. He identified two major problems. First, well-learned information was catastrophically forgotten as new information was learned, in both small and large networks. Second, even a single learning trial with new information caused significant loss of old information, paralleling McCloskey and Cohen's results.
Ratcliff also observed that outputs were often a blend of previous and new inputs. In larger networks, items learned in groups (e.g., AB then CD) were more resistant to forgetting than items learned individually (e.g., A then B then C), though forgetting remained substantial. Adding new hidden units did not reduce interference.
Implications for Modern AI
Catastrophic interference remains a significant challenge in contemporary Artificial intelligence systems, including Large language models and Transformer (architecture)-based architectures. When these models are fine-tuned on new tasks or datasets, they can lose performance on previously learned tasks. This issue is particularly relevant for continual learning, where models must adapt to new information without forgetting old knowledge.
Researchers have developed various strategies to mitigate catastrophic forgetting, such as Curriculum Learning, Model Pruning, and regularization techniques. However, no universal solution exists, and the problem remains an active area of research in Neural network theory and application.
Conclusion
Catastrophic interference highlights a fundamental limitation of standard neural networks and underscores the complexity of replicating human memory. The work of McCloskey and Cohen and Ratcliff established the phenomenon and its characteristics, providing a foundation for ongoing efforts to build more stable and adaptable learning systems.