Fig. 9: Quality and diversity of generated samples.
From: Brain-inspired replay for continual learning with artificial neural networks
For the class-incremental learning scenario on CIFAR-100, compared are the quality and diversity of the generated samples that are replayed by standard generative replay (GR) with individual modifications added ('+', left within each panel) and by brain-inspired replay (BI-R) with individual modifications removed ('−', right). All measures are computed for samples generated by the model after it was incrementally trained on all 100 classes; for the measures in the last two panels those generated samples were compared to real samples from the test set. a Our modified version of Inception Score (Modified IS). Higher means better samples. b Our modified version of Frechet Inception Distance (Modified FID). Lower means better samples. c Precision & Recall curves. Towards the top indicates better sample quality (or ‘precision'), towards the right indicates better sample diversity (or ‘recall'). Each bar or thick curve reflects the mean over 10 repetitions, error bars are ±1 SEM, individual repetitions are indicated by dots or thin curves. rtf replay-through-feedback, con conditional replay, gat gating based on internal context, int internal replay, dis distillation.
