In the context of continual incremental learning, the two most discussed challenges are concept drift and catastrophic forgetting. I am looking for recourses that discuss each of these problems separately. Also I am wondering about any research efforts that discuss the two challenges together. How they are related and affected by each other?
Dear Alia Alkameli
Continual learning algorithms attempt to replicate this ability in neural networks and tackle the catastrophic forgetting problem. In essence, continuous learning is the progressive acquisition of new activities.
According to Robins (1995), rehearsing techniques can avoid catastrophic forgetting. This implies that when new data is introduced, the neural network gets retrained on some of the previously learnt data. However, previously learned knowledge may not be available for such retraining in general.
Standard neural network topologies, on the other hand, suffer from catastrophic forgetting, making it impossible for them to learn a succession of tasks. Without resolving this issue, a NN will struggle to adapt to lifelong or continuous learning, which is critical for AI.
Kind Regards
Qamar Ul Islam
Alia Alkameli ,
When one tries to compensate for concept drift catastrophic forgetting may occur. That is, machine learning assumes that a training set is used that is a representative sample of the underlying unknown distribution. If there is a shift on the distribution space, a new modelling including this shift should be included by expanding the distribution space. If the initial space is just shifted instead of adjusting the size of the distribution space, then catastrophic forgetting will occur.
One way, regarding catastrophic forgetting, may be found here: https://inspirehep.net/literature/295249 ``Retrieval properties of neural networks with infinitely many marked patterns''.
- Badges
- Science topic