Rethinking hardware architecture for probabilities and sampling
Focusing on sampling method
Many probabilistic problems are intractable, in the sense that the exact inference cannot be
computed in a reasonable time. This is the case in particular when there is no analytical
solution (as opposed to a Kalman filters, for example) or when the possible values of unknown
variables are too large to be enumerated.
Variational inference methods aim to find an optimal solution to a problem that approximates the initial problem, unlike Markov-Chain Monte-Carlo (MCMC) methods which aim to find an approximate solution to the initial exact problem.
Among MCMC methods, gradient-based algorithms are efficiently accelerated on GPU but tend to get stuck in local minima. Sampling-based methods are very promising from a mathematical standpoint but are particularly challenging for existing hardware.
Our key advantage for sampling
We developed and patented an architecture that allows to perform sampling operations at an unmatched speed. Where classical architectures must compute explicitly products of probability distributions (very long vectors) before selecting one sample, our architecture allows to evaluate in parallel an approximation of the all the products and manage to select a sample faster and with less energy.
Optimizing the architecture at all levels
Sampling is not the only aspect in which our architecture
is dedicated for probabilistic computing
Compression of probability distributions to reduce the model size.
High-performance scalar and vector units for parameter estimations.
Low-overhead synchronization between circuit parts.