Time: Friday, October 12th, 09:00
Speaker: Kerem CAMSARI, Purdue University
Digital computers are built out of bits that represent information in deterministic states 0 and 1. At the other end, quantum computers are built out of q-bits that represent information as delicate superpositions of 0 and 1. In this talk, I will introduce our body of work on probabilistic bits (p-bit) that randomly fluctuate between 0 and 1, placing p-bits conceptually in between deterministic and quantum bits. We have shown that hardware p-bits can be realized with present day, room temperature devices using magnetic and non-magnetic components and p-circuits built of interconnected p-bits can be useful for a wide variety of real world applications that are inherently probabilistic, just as quantum bits are naturally suited for inherently quantum mechanical problems.
We have shown that p-circuits can be broadly relevant for Quantum Computing and Machine Learning inspired problems. In the context of Machine Learning, p-bits can function as low-level representations of Binary Stochastic Neurons (BSN), therefore they can be used to realize efficient hardware implementations for Bayesian and Inference Networks, as well as hardware accelerators for BSN-based statistical learning algorithms. In the context of Quantum Computing, p-bits can be used for optimization problems, covering both classical and quantum annealing, to solve hard problems such as Traveling Salesman and Integer Factorization. Inverse problems such as Integer Factorization are enabled by the "invertibility" feature of p-circuits, where a logic gate not only finds the appropriate output for a set of fixed inputs but also the appropriate inputs for a fixed set of outputs. I will illustrate each application by a representative example either by using experimentally benchmarked device models simulated in circuit simulators such as SPICE or by hardware implementations of p-circuits using non-magnetic building blocks.