cluster-mil is a weakly supervised learning technique which uses concepts from the Multiple Instance Learning (MIL) framework to train a convolutional deep neural network. As data is only available in batches, we pretrain a variational autoencoder (unsupervised) and then estimate class labels from weak labels provided during training.
This work was presented at ML4H @ NeurIPS 2018 and a full description of the method is available on ArXiv: Cluster-Based Learning from Weakly Labeled Bags in Digital Pathology
The implementation of the MNIST-BAG experiment is given in mnist-bag.py and utilizes our novel loss function (loss.py). Pretrained autoencoder models are also provided in /latent_models.