Learning Self-train for semi-supervised few shot classification

NeurIPS 19 11/3/2020

Motivation

This paper is not hard to understand, the slide: (https://drive.google.com/file/d/151ZyvJK77nPJ36LA2gdk3S--8caXS43-/view) has been very clear.

Notations:

• $\Phi_{ss}$ : feature extractor of base learner, one of meta-parameters

• $\theta$ : final layer classifier of base learner

• $\theta'$ : initialization parameters of $\theta$ , one of meta-parameters

• $\Phi_{swn}$ : weights of soft-weighting network, one of meta-parameters

Inner loop:

1. Pseudo-labeling

2. Cherry-picking: (hard selection, soft weighting)

3. Self-training: re-training (S+R), fine-tuning (S)

Outer loop:

1. update $\Phi_{swn}$ after re-training using the validation loss on query set based on $\theta_m$

2. update $[\Phi_{ss}, \theta']$ after fine-tuning using the validation loss on query set based on $\theta_T$

Reference

• https://papers.nips.cc/paper/9216-learning-to-self-train-for-semi-supervised-few-shot-classification.pdf

• https://github.com/xinzheli1217/learning-to-self-train

• https://drive.google.com/file/d/151ZyvJK77nPJ36LA2gdk3S--8caXS43-/view

• https://qianrusun.com/