Mitigating Catastrophic Forgetting in Continual Learning-Based Image Classification
Swarali Borde · 2023
The report investigates catastrophic forgetting (CF) in the context of continual learning using neural networks for image classification tasks. The study employs simple and small networks on datasets like the MNIST and split-CIFAR10, observing the immediate loss of knowledge when networks switch jobs. While findings align with some existing literature, severe forgetting is attributed to increased jobs, smaller networks, and lack of interleaved training. Notably, a more extensive network with a pretraining dataset exhibits more gradual forgetting, suggesting potential mitigation. A technique involving separate classifiers for each task is proposed, though practicality hinges on task awareness and classifier availability. The report recommends further exploration, suggesting experiments with more extensive networks pre-trained on diverse datasets and hybrid latent representations from classification and reconstruction models. The potential of the VAE is highlighted, as is empirically analyzing forgetting across different task numbers in split datasets.