SemGIF: A Semantics Guided Incremental Few-shot Learning Framework with Generative Replay

S. Divakar Bhat, Biplab Banerjee, Subhasis Chaudhuri
2021 British Machine Vision Conference  
We address the problem of incremental few-shot learning (IFSL) by leveraging the notion of generative feature replay. Learning novel concepts while preserving old knowledge is a long-lasting challenge in machine learning. The main concern in IFSL is to combat the catastrophic forgetting of the base classes whose training data are not available during the incremental stage while ensuring good generalization for the few-shot classes. Existing techniques prefer to preserve some base class samples
more » ... o tackle forgetting, which does not comply with the intention of incremental learning. To this end, we propose a novel framework called Semantics Guided IFSL (SemGIF), which trains a generative model to synthesize base class samples on demand during the incremental step. Considering the importance of modeling a discriminative feature space in IFSL for separating the base and the novel classes, we propose a feature augmentation strategy where the visual embeddings are supplemented with the semantic features obtained from a word-embedding space. Such a feature space is found to produce enriched class prototypes to be utilized during classification. Experimental results on CIFAR-100, CUB, mini-ImageNet, and tiered-ImageNet in the homogeneous (within-dataset) and a novel heterogeneous (cross-dataset) setup showcase sharp improvements than the literature.
dblp:conf/bmvc/BhatBC21 fatcat:fvh2rivqbrbcjjdiqkpbt5hdzu