Reproduce “Dynamic Few-Shot Visual Learning without Forgetting” on OpenResearch

About this paper

Spyros Gidaris, Nikos Komodakis

Researchers at University Paris-Est developed a few-shot visual learning system capable of efficiently learning new categories from minimal examples at test time, crucially without compromising recognition accuracy on previously learned base categories. This framework, utilizing a cosine-similarity based classifier and an attention-based weight generator, achieved state-of-the-art performance on Mini-ImageNet and other benchmarks, notably maintaining base category accuracy around 70.50%.

Key insights

• The cosine-similarity based classifier fosters the learning of more discriminative and generalized feature representations, crucial for recognizing unseen categories and unifying the recognition process.
• Explicitly leveraging "past knowledge" from base category weights through an attention mechanism in the weight generator significantly enhances the system's ability to learn novel categories from very few examples, particularly in the 1-shot setting.
• Removing the ReLU non-linearity from the last hidden layer of the feature extractor further improves novel category recognition when combined with the cosine-similarity classifier.

Summary by alphaXiv.