In this paper we combine uncertainty-biased molecular dynamics with active learning to show how we can learn machine learning interatomic potential (MLIP) models that are more robust to predictions on extrapolative regions.

We propose a temporal knowledge graph baseline based on the simple principle of recurrency. Surprisingly, a simple approach like this beats state of the art methods based on representation learning on 3 out of 5 commonly used datasets!

We propose a probabilistic model that bridges graph machine learning and sum-product networks to tractably answer probabilistic queries. Our Graph-induced Sum-Product Network can solve unsupervised and supervised graph tasks and it is especially effective in modeling the missing data distribution as well as exploiting unlabeled data in a scarce supervision scenario.

In this paper I try to address the following question: “Is it really worth to build a k-Nearest Neighbor graph from tabular data and then apply deep graph nets on top of it to classify each sample in the table?

In this paper we accelerate HMC simulations by learning atom-dependent timesteps and the number of molecular dynamics steps! With an appropriate choice of the loss, we can significantly speed up simulations and reduce autocorrelation times by 25% on alanine dipeptide.