Deep Reference Priors (ICML 2022) pre-train the weights of a neural network with unlabeled data, using the theory of reference priors. It allows the data to dominate the posterior rather than the choice of the prior. The prior is represented by a finite number of particles (or neural networks) which are trained to span the prediction space.
This method is competitive with other SoTA semi-supervised learning methods.
conda create -n ref_prior --file setup.yml
conda activate ref_prior
python train.pyParts of this code are adapted from https://github.com/kekmodel/FixMatch-pytorch