Active self-semi-supervised learning for few labeled samples

Key observations:

1. Weight initialization struggles to effectively transfer valuable information obtained from self-supervised training to the semi-supervised model
2. Label propagation on pre-trained features to construct prior pseudo-labels, serving as an effective intermediary to transfer information from self-supervised learning to semi-supervised models.
3. Adapt active learning to improve the accuracy of prior pseudo-labels.

Recommended citation: Ziting Wen, Oscar Pizarro, and Stefan Williams. "Active self-semi-supervised learning for few labeled samples." Neurocomputing (2024): 128772.
Download Paper