Self-supervised learning
Reinhard Heckel · 2025
Abstract Supervised training requires pairs of target image and associated measurements, which are often difficult to collect. This chapter discusses self-supervised learning approaches based on constructing a self-supervised loss for training a neural network to map a measurement to a clean image. Three different approaches are discussed: The first are self-supervised loss functions that rely on two independent noisy measurements of the same signal and enable the construction of unbiased estimates of the supervised loss. The second are self-supervised loss functions based on Stein’s unbiased estimator, and the third are masking-based approaches.