Boosting Representation Learning via Similarity-based Active Data Sampling from the Web

Training deep learning models requires a huge dataset, and human annotation is necessary to construct dataset. However, the larger the dataset, the greater the annotation cost. Self-Supervised Learning (SSL), a form of representation learning, addresses this problem by using mechanically generated pseudo-labels for training. SSL eliminates the need for human annotation, thereby reducing the costs associated with dataset construction and facilitating the training on large-scale dataset, compared to supervised learning. However, when it is challenging to collect a large number of images due to varying capture conditions, the dataset size diminishes, and SSL performance degrades. This study proposes a method that combines image similarity search with a diversity-based query strategy for Deep Active Learning, selectively augmenting the training dataset with images collected from the web. Our approach enables the collection of images highly relevant to downstream tasks via similarity search, while excluding images that are too similar to those in the training dataset, thereby maintaining dataset diversity with a diversity-based query strategy. Experiments demonstrate that our method enhances SSL performance, particularly when the training dataset for downstream tasks is limited.