Shaila's work on Deep Boltzmann Machines out in Nature Electronics!

Can Ising Machines be useful for generative AI, beyond their typical use for solving combinatorial optimization problems?

We tackle this question in our latest work just published in Nature Electronics led by Shaila Niazi:

https://rdcu.be/dK2gd

1. We show that energy-based *deep* Boltzmann machines (BM), traditionally considered to be intractable, can be trained using dedicated probabilistic computers where up to ~50 billion probabilistic samples per second are taken.

2. We find that the deep but sparse BMs with only 30,000 parameters can reach beyond the capabilities of fully-connected Restricted BMs with 3,000,000 parameters, where sparse BMs can generate new images where RBMS fail.

3. We train the full MNIST, Fashion MNIST and CIFAR datasets which shows promising classification accuracy "out-of-the-box" without any algorithmic fine-tuning.

We are very excited about the potential of p-bit based Ising Machines (via their digital as well as scaled analog realization) on generative, physics-based AI algorithms.

We acknowledge a fruitful collaboration with Yao Qin, Masoud Mohseni and grateful to Office of Naval Research for funding this work.