Stochastic Systems Group  

Stochastic Logic Circuits for Inductive Inference
Vikash K. Mansinghka
MIT
Inductive inferences, formalized via queries on probability models, are growing increasingly important in machine learning, cognitive science and signal processing. Currently, algorithms for performing these inferences are executed via conversion to real arithmetic, which is simulated to finite precision using floatingpoint circuits built out of combinational boolean logic. This approach results in several inefficiencies, chiefly the computation of many unwanted (and ultimately implausible) bits of answers to queries, and significant silicon/cost overhead for exploiting the finegrained parallelism inherent in many Monte Carlo approaches to inference.
Motivated by these difficulties, we introduce combinational stochastic logic, an extension of combinational boolean logic that is closed under expressive means of combination and abstraction. Stochastic combinational logic naturally supports the construction of parallel circuits that perform Monte Carlo computations, and has several interesting theoretical and practical properties. We will also present the results of preliminary VHDL simulations of a circuit for Gibbs sampling an Ising model (with conditional probabilities represented to 8 bits), that can perform roughly 8 million Gibbs sweeps of a lattice per second when clocked at 100MHz, independent of the size of the model, with linear space costs.
Problems with this site should be emailed to jonesb@mit.edu