OpenIncrement: A Unified Framework for Open Set Recognition and Deep Class-Incremental Learning

Jiawen Xu,Claas Grohnfeldt,Odej Kao

2023-10-06

Abstract:In most works on deep incremental learning research, it is assumed that novel samples are pre-identified for neural network retraining. However, practical deep classifiers often misidentify these samples, leading to erroneous predictions. Such misclassifications can degrade model performance. Techniques like open set recognition offer a means to detect these novel samples, representing a significant area in the machine learning domain. In this paper, we introduce a deep class-incremental learning framework integrated with open set recognition. Our approach refines class-incrementally learned features to adapt them for distance-based open set recognition. Experimental results validate that our method outperforms state-of-the-art incremental learning techniques and exhibits superior performance in open set recognition compared to baseline methods.

Computer Vision and Pattern Recognition,Machine Learning

What problem does this paper attempt to address?

The paper aims to address the problem of how neural networks learn representations suitable for Open Set Recognition (OSR) in the context of class-incremental learning. Specifically, the paper focuses on:

1. **Feature Distortion Problem**: During the class-incremental learning process, due to the phenomenon of catastrophic forgetting, the feature space gets distorted, causing features of the same class to no longer cluster tightly, and features of different classes to potentially intertwine. This distortion undermines the two fundamental assumptions required for open set recognition:
- Data representations of the same class should be close to each other (Assumption 1);
- Data representations of different classes should be far apart from each other (Assumption 2).

2. **Integrating OSR with Class-Incremental Learning**: Most current class-incremental learning methods require pre-identification of new samples when dealing with new classes, which often leads to misclassification in practical applications. Therefore, the paper proposes a unified framework that combines open set recognition with class-incremental learning to address the above issues and improve the robustness and accuracy of the model when new classes appear.

Through experimental validation, this method achieves significantly better results than existing techniques on public continual learning benchmarks and also performs excellently in open set recognition.