Oubli catastrophique is a phenomenon observed in intelligence artificielle, particularly in apprentissage automatique models, where a model forgets previously learned information upon learning nouvelles données or tasks. This issue is especially prominent in neural networks that are trained sequentially on different tasks, leading to a decline in performance on earlier tasks.
Le terme a été décrit pour la première fois dans le contexte de la psychologie théorie de l'apprentissage, but it has significant implications in AI and machine learning. When a model is trained on one task and then retrained on a different task, it often fails to retain the knowledge acquired from the first task, resulting in poor performance on it. This is problematic for applications where continuous learning and adaptation are required.
Catastrophic forgetting is commonly attributed to the way neural networks update their weights during training. When new data is introduced, the adjustments made to the network’s parameters can interfere with the previously established representations, causing them to degrade or be overwritten. As a result, the model exhibits a high degree of sensitivity to the order in which tasks are presented.
Pour atténuer l'oubli catastrophique, les chercheurs ont exploré diverses techniques, notamment :
- Régularisation méthodes : These help maintain the weights associated with previous tasks while allowing new learning.
- Réseaux augmentés par la mémoire : These incorporate la mémoire externe structures pour stocker les informations des tâches précédentes.
- Réseaux neuronaux progressifs : These architectures allocate separate resources for each task, preserving knowledge from earlier tasks.
Comprendre et traiter l'oubli catastrophique est crucial pour le développement systèmes d'IA that can learn continuously and adapt to new information without losing valuable knowledge.