2020 Interesting papers and ideas

Learning To Classify Images Without Labels

• PDF
• My implementation of the core idea
• Core idea: Learn (or use pre-trained model) unsupervised representations. Then use k-means to cluster images. For each image, take nearest n(=5ish) neighbors and train a model to classify. You want to maximize dot product of softmax out + entropy of label predictions.

Rethinking the Truly Unsupervised Image-to-Image Translation

• PDF
• Code

Implicit Neural Representations with Periodic Activation Functions

• PDF
• This gentleman will hopefully make the code open source soon

Bootstrap Your Own Latent A New Approach to Self-Supervised Learning

• PDF
  • Built on this paper - Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results
• Thoughts: interesting idea but a long road to it being applied for non-research use.