Dynamic Curriculum Learning via In-Domain Uncertainty for Medical Image Classification

Chaoyi Li, Meng Li, Can Peng, Brian C. Lovell

Abstract: This paper presents an innovative approach to curriculum learning, which is a technique used to train learning models. Curriculum learning is inspired by the way humans learn, starting with simple examples and gradually progressing to more challenging ones. There are currently two main types of curriculum learning: fixed curriculum generated by transfer learning, and self-paced learning based on loss functions. However, these methods have limitations that can hinder their effectiveness. To overcome these limitations, this article proposes a new approach called Dynamic Curriculum Learning via In-Domain Uncertainty (DCLU), which is derived from uncertainty estimation. The proposed approach utilizes a Dirichlet distribution classifier to obtain prediction and uncertainty estimates from the network, which can be used as a metric to quantify the difficulty level of the data. An uncertainty-aware sampling pacing function is also introduced to adapt the curriculum according to the difficulty metric. This new approach has been evaluated on two medical image datasets, and the results show that it outperforms other curriculum learning methods. The source code for this approach will be released at https://github.com/Joey2117/DCLU.