Title: Hardware Accelerated Restricted Boltzmann Machines for Computationally Difficult Problems
Presenter: Saavan Patel (UC Berkeley)
Abstract: Stochastic formulations of the Ising Problem, in the form of the Boltzmann Machine and Restricted Boltzmann Machine (RBM) can solve NP-Hard and NP-Complete combinatorial optimization problems. These systems can be trained using various Machine Learning algorithms to model arbitrary probability distributions, and solve problems in a stochastic manner. We explore a new hardware paradigm based on Markov Chain Monte Carlo techniques and probabilistic transitions taken directly in hardware. We show a methodology and a general purpose hardware implementation in both digital and mixed-signal CMOS for combinatorial optimization and integer factorization problems. In addition, we propose novel algorithmic methods of merging models to solve larger problems and ease the convergence of the training algorithm. Using this combined hardware-software approach, we demonstrate factorization of 20 bit numbers in software, and hardware accelerated factorization of 16 bit numbers. Our results further provide a pathway for scalable hardware implementation of combinatorial optimization problems such as the Traveling Salesman Problem, Max Cut etc, along with integer factorization problems and possible applications in quantum mechanical simulations.
Bio: Saavan Patel is a doctoral student in professor Sayeef Salahuddin's group studying novel methods of computation and non von Neumann architectures. His work spans fields from machine learning, through algorithm design, analog and digital circuits, and device physics. He graduated from the University of California, Berkeley with a BS in Electrical Engineering and Computer Science.
Please note: This e-Workshop is only available to the JUMP research community, such as Principal Investigators, postdoc researchers, students, and corporate sponsors. ASCENT is one of six JUMP centers administered by SRC. For access to full program information, please go to src.org. Thank you for interest in ASCENT.