Restricted Boltzmann Machines revisited: from landscape to transition path sampling
Remi Monasson (Cnrs et Ecole Normale Supérieure)
Hybrid seminar: onsite + zoom.
https://cnrs.zoom.us/j/93762609075?pwd=eXhrNUQ3clhDNGx6dHpBQjVIdXVkZz09
Meeting ID: 937 6260 9075
Passcode: p5SKcT
Restricted Boltzmann Machines (RBM) were introduced thirty-five years ago in the context of unsupervised learning of data distributions. The simplicity of their architecture make RBM both potentially interpretable and amenable to analytical treatment with statistical physics methods. I will show how these methods can help understand « good » operation regimes for RBMs, in particular how to stack RBMs to speed up sampling, an idea first introduced under the name of deep tempering by Bengio and collaborators. I will also briefly discuss recent applications to transition path sampling for protein design.