Icml2022 | galaxy: active learning of polarization map

Thesis link ：https://arxiv.org/pdf/2202.01402.pdf

Active learning is an efficient method of marking , By interactively selecting a small part of unlabeled data for marking and training , So as to train an efficient model . stay “ Open world ” Setting up , The classes of interest make up only a small part of the entire data set —— Most data may be treated as non distributed or unrelated classes . This leads to extreme category imbalances , Our theories and methods focus on this core problem . We propose a new active learning strategy , be called GALAXY (Graph-based active learning At the eXtrEme), It combines the idea of graph based active learning and deep learning . Compared with most active learning methods ,GALAXY It can automatically and adaptively select more category balanced samples for marking . Our theory shows that ,GALAXY A fine sampling of uncertainty is performed , It collects a more class balanced data set than ordinary uncertainty sampling . Through the experiment , We demonstrate that in unbalanced visual classification settings generated from popular data sets ,GALAXY It is superior to the existing advanced deep active learning algorithm .