Reproduce “Partner-Assisted Learning for Few-Shot Image Classification” on OpenResearch

About this paper

Jiawei Ma, Hanchen Xie, Guangxing Han, Shih-Fu Chang, Aram Galstyan, Wael Abd-Almageed

The Partner-Assisted Learning (PAL) framework introduces a two-stage training method for few-shot image classification, utilizing a "Partner Encoder" trained with supervised contrastive learning to provide robust, instance-level regularization for a "Main Encoder." This approach consistently achieves state-of-the-art accuracy across four benchmark datasets, improving 1-shot classification on miniImageNet by 4.55% over previous methods.

Key insights

• Decoupling the learning of robust, instance-level "soft-anchors" from the main classification feature extraction process provides superior generalization for few-shot tasks.
• Supervised contrastive learning, when used to train the Partner Encoder, proves to be a more effective source of regularization for the Main Encoder than unsupervised contrastive learning or direct cross-entropy.
• The proposed logit and feature-level alignment losses, which leverage the Partner Encoder's instance-level similarity, are more effective than traditional KL-divergence distillation, which attempts to minimize the negative entropy of the Partner Encoder's output.

Summary by alphaXiv.