Timezone: »

The IID: A Natively Probabilistic Reconfigurable Computer
Vikash Mansinghka

Tue Dec 08 12:00 AM -- 12:00 AM (PST) @ None

We demonstrate the IID, a natively probabilistic, reconfigurable digital computer based on stochastic digital circuits. The IID Mark 0 is implemented on top of a commodity array of Field Programmable Gate Arrays (FPGAs), and programmed using a generic toolchain based on the State-Density-Kernel abstractions from Blaise. It can be used to perform MCMC inference on factor graphs with hundreds of thousands of variables in real time. The technical novelty of the IID rests in the stochastic digital circuits which it simulates. These circuits enable the construction of massively parallel, low bit precision, fault tolerant machines that directly simulate Markov chains in hardware, leveraging small XOR-SHIFT RNGs to provide the necessary stochasticity. Due to space constraints, we refer the reader to FIXME cite for technical details about these circuits, including a detailed discussion of their capabilities and their novelty and an overview of the software tools used to program the IID to solve arbitrary discrete factor graphs. The key consequence of our approach is that the IID can be used to solve probabilistic graphical model problems with 2-3 orders of magnitude improvement in price/power/performance product and 3-6 orders of magnitude improvements in robustness to bit errors than is possible using conventional computer architectures.

Author Information

Vikash Mansinghka (Massachusetts Institute of Technology)

Vikash Mansinghka is a research scientist at MIT, where he leads the Probabilistic Computing Project. Vikash holds S.B. degrees in Mathematics and in Computer Science from MIT, as well as an M.Eng. in Computer Science and a PhD in Computation. He also held graduate fellowships from the National Science Foundation and MIT’s Lincoln Laboratory. His PhD dissertation on natively probabilistic computation won the MIT George M. Sprowls dissertation award in computer science, and his research on the Picture probabilistic programming language won an award at CVPR. He served on DARPA’s Information Science and Technology advisory board from 2010-2012, and currently serves on the editorial boards for the Journal of Machine Learning Research and the journal Statistics and Computation. He was an advisor to Google DeepMind and has co-founded two AI-related startups, one acquired and one currently operational.

More from the Same Authors