Predictive Information Bottleneck is an information theory method to learn a low-dimensional representation of a system that has maximum predictive power.
The Predictive Information Bottleneck (PIB) is similar to the reweighted autoencoded variational Bayes for enhanced sampling. Both methods use iterations between molecular dynamics simulations and deep learning to learn a probability distribution along a latent space and a physically interpretable progress coordinate to enhance sampling. PIB is used to learn the progress coordinate which is maximally predictive of the trajectory’s future behavior.
For examples of previously performed studies in which Predictive Information Bottleneck was the primary method used, see the following example cases: