Dataset Bias Prediction for Few-Shot Image Classification

Jangwook Kim, Kyung-Ah Sohn

Keywords: Machine Learning, Few-Shot Learning, Image Classification, Dataset Bias

Abstract: One of the obstacles which negatively affect the image classification performance is dataset bias. In particular, if each class has only a few training data samples, the data are highly likely to have dataset bias. Therefore, dataset bias can be a serious issue in few-shot learning, but has rarely been studied so far. To address this issue, we propose a bias prediction network to help improve the performance of few-shot image classification models. Once the features are extracted from an image data, the bias prediction network tries to recover the bias of the raw image such as color from the features. However, if the bias prediction network can recover it easily, we can assume that the extracted features also contain the color bias. Therefore, in our proposed framework, the full model tries to extract features that are difficult for the bias prediction network to recover from. We validate our method by adding the bias prediction network to several existing models and evaluating the performance improvement. Our experimental results show that the bias prediction network can suppress the negative effect of the dataset color bias, resulting in the substantial improvements in existing few-shot classification models. The proposed bias prediction network, which can be integrated with other models very easily, could potentially benefit many existing models for various tasks.

One-sentence Summary: The bias prediction network improves the performance of the original few-shot classification model by minimize the bias of datasets.