Unsupervised Confidence Calibration for Active Learning

概要

In active learning, training data to be annotated is generally selected based on the confidence scores output by a classifier. However, the selection of the initial training data pool typically relies on random sampling since the classifier has not yet been trained. To address this issue, it is necessary to obtain a classifier that can appropriately assess confidence without supervised learning at the initial pool selection stage. In this study, we propose an unsupervised confidence calibration method based on domain adaptation. The proposed framework integrates a variational autoencoder into an unsupervised domain adaptation method and trains them simultaneously, thereby allowing the estimation of the joint distribution of input data and class labels. The proposed framework enables the classifier to compute reliable confidence scores for the target domain in a completely unsupervised manner. We conducted experiments on medical image datasets to evaluate the effectiveness of the proposed unsupervised confidence calibration method and its impact on active learning. Experimental results showed that the proposed method consistently outperforms the baseline in all six transfer settings, achieving an average improvement of 1.0–1.5 points in classification accuracy across domains.

福井 直哉

博士前期課程学生

早志英朗

准教授

深層学習やベイズ推定を基盤とした機械学習アルゴリズムの開発を中心に、生体信号解析、医用画像処理などの応用研究に従事。

長原一

教授

コンピューテーショナルフォトグラフィ、コンピュータビジョンを専門とし実世界センシングや情報処理技術、画像認識技術の研究を行う。さらに、画像センシングにとどまらず様々なセンサに拡張したコンピュテーショナルセンシング手法の開発や高次元で冗長な実世界ビッグデータから意味のある情報を計測するスパースセンシングへの転換を目指す。