1. Representation learning: Pick a pre-training/pre-text task for learning inherent representations in images that are not just low-level features. Here they us

1. Representation learning: Pick a pre-training/pre-text task for learning inherent representations in images that are not just low-level features. Here they use augmentations on images for such task to minimize the loss between real and augmented images. This forces the network to learn high-level features important for discrimination of these two images. 2. Self-clustering: Samples are drawn from neighborhood of embedded learned space to assign cluster labels. They push the network to make neighborhood samples to same cluster by including this in SCAN loss. 3. Fine tuning: Assign labels to clusters specific to your data.