Lukas Helmig

---

I will introduce Restricted Boltzmann Machines (RBM) from a statistical physics perspective. Practical aspects such as training and use-cases as well as the relevance of RBMs in modern machine learning will be discussed. A mapping to Hopfield networks will be presented that allows to treat RBMs (or more specifically Binary-Gaussian RBMs) as an associative memory. Similarly to the Hopfield model, the RBM as an associative memory has a critical pattern storage capacity, which will be derived. A connection of this capacity to the training of RBMs will be illustrated. I will end by sketching future lines of research such as a potential Gardner-type learning algorithm for RBMs, interesting Hopfield models to investigate and a connection of the Hopfield-type capacity to the expressivity of (Quantum) RBMs.

---

RWTH Aachen
Zoom join online
Contact: Sebastian Diehl