iSSL-AL: a deep active learning framework based on self-supervised learning for image classification

• Published in: Neural computing & applications Vol. 36; no. 28; pp. 17699 - 17713
• Main Authors: Agha, Rand, Mustafa, Ahmad M., Abuein, Qusai
• Format: Journal Article
• Language: English
• Published: London Springer London 01-10-2024; Springer Nature B.V
• Subjects: Algorithms; Annotations; Artificial Intelligence; Artificial neural networks; Classification; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Datasets; Deep learning; Image annotation; Image classification; Image Processing and Computer Vision; Image retrieval; Labeling; Machine learning; Medical research; Original Article; Probability and Statistics in Computer Science; Sampling; Self-supervised learning; Uncertainty; Deep learning; Active learning; Self-supervised learning
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-024-10271-6

Deep neural networks have demonstrated exceptional performance across numerous applications. However, DNNs require large amounts of labeled data to avoid overfitting. Unfortunately, the labeled data may not be available; annotating large amounts of data is time-consuming, laborious, and requires human expertise, making it unfeasible to rely on manpower for annotation. One of the solutions to address this limitation is active learning (AL), a technique that utilizes unlabeled data while maintaining high performance. AL reduces the annotation cost of large datasets and enhances the training of models with fewer annotations. Uncertainty sampling has been proven to be one of the most effective strategies in AL; however, it lacks diversity. This research proposes iSSL-AL, a novel active learning framework that utilizes self-supervised learning (SSL) to ensure informative yet diverse samples. Three main aspects categorize the novelty of our work. The first is extending the margin uncertainty sampling by incorporating SSL to select informative and diverse points. The second is employing incremental learning for efficient training of the AL base classifier, where the model is trained incrementally in each AL cycle. The third is addressing the cold start problem, as our framework achieved high results in the early stages of training. Experiments show that iSSL-AL outperforms other state-of-the-art algorithms on the MNIST, FashionMNIST, and CIFAR-10 datasets, with accuracy scores of 99%, 98.9%, and 93.5%, respectively, effectively selecting diverse and informative samples.