Online Generative Replay for Online Class-Incremental Continual Learning

Sammanfattning

Continual learning in deep neural networks remains a critical chal-lenge due to catastrophic forgetting, particularly in online class incre-mental learning, where the data distribution changes over time andthe model must learn sequentially without revisiting the full pastdataset. Experience Replay is a widely used baseline that alleviatesforgetting by storing real samples in a buffer; however, it raises mem-ory and privacy concerns. This thesis explores whether generativereplay strategies, specifically Deep Generative Replay, can match oroutperform Experience Replay under strict online continual learningconstraints, where training is performed in a single pass over thedata stream and revisiting past information is possible only throughreplay.We propose Online Generative Replay with buffer, a hybrid replaystrategy based on online generative replay that combines a class-conditional variational autoencoder with a small episodic memoryof real samples. Online Generative Replay with buffer replays bothgenerated samples and buffered real examples during online learn-ing. We evaluate Online Generative Replay with buffer on CIFAR-100 and Mini-ImageNet under an online class-incremental protocoland compare it against Experience Replay and online generative re-play without a real buffer. Our experimental analysis reports finalaccuracy across buffer sizes and quantifies retention and compute us-ing end-forgetting and total training time, highlighting the regimeswhere hybrid replay is most effective.Experiments on two standard online class-incremental benchmarks,CIFAR-100 and Mini-ImageNet, show that Online Generative Replaywith buffer consistently improves over online generative replay with-out a real buffer and is competitive with Experience Replay, particu-larly in memory-constrained regimes (for example, buffer size 250).Across both datasets, Online Generative Replay with buffer achievesstronger retention than Experience Replay in terms of end-forgetting,while Experience Replay can remain faster in total training time. Inaddition, task-wise visualizations indicate that Online Generative Re-play with buffer preserves earlier-task performance more reliably bythe end of the stream, reflecting reduced forgetting as training pro-gresses.These results suggest that hybrid generative replay such as On-line Generative Replay with buffer can be a competitive alternativeto buffer-only rehearsal in online continual learning, particularly inmemory-limited regimes (for example, buffer size 250) where it matchesor exceeds Experience Replay in final accuracy and remains consistently stronger than generator-only replay. In addition, the reportedend-forgetting results on Mini-ImageNet indicate that Online Gener-ative Replay with buffer can improve retention compared to Experi-ence Replay, at the cost of increased training time due to online gen-erator updates. We also identify several directions for future work,including using selected buffered exemplars to further stabilize gen-erator training, dynamically adapting the generator update schedule,and scaling to higher-resolution datasets using stronger generativearchitectures.