Probabilistic Machines for Low-level Sensor Interpretation



We have a fully funded Postdoctoral Fellowship available for 2017-2019 : (see the overview and the project presentation )


Scientific Context

The development of modern computers is mainly based on increase of performances and decrease of size and energy consumption. This incremental evolution is notable, but it involves no notable modification of the basic principles of computation. In particular, all the components perform deterministic and exact operations on sets of binary signals. These constraints obviously impede further sizable progresses in terms of speed, miniaturization and power consumption. As detailed below, the goal of the MicroBayes project is twofold:

  1. to investigate a radically different approach to perform computations, namely stochastic computing using stochastic bit streams.
  2. to show that stochastic architectures can outperform standard computers to solve complex inference problems both in terms of execution speed and of power consumption.

We will evaluate stochastic machines on difficult Bayesian inference problems. Moreover we will demonstrate the interest and feasibility of stochastic computing on two applications involving low-level information processing from sensor signals. The given application are sound source localization and separation as shown below:

  • Source localization:


Source localization

  • Source separation:



  • Emmanuel Mazer (DR CNRS LIG) emmanuel.mazer imag.fr
  • Laurent Girin (Prof. Grenoble-INP GIPSA-lab) laurent.girin inria.fr
  • Laurent Fesquet (MCF Grenoble-INP TIMA) laurent.fesquet imag.fr
  • Didier Piau (Prof. UJF Institut Fourier) didier.piau univ-grenoble-alpes.fr
  • Raphael Frisch (Phd with a grant from Persyval Lab) raphael.frisch univ-grenoble-alpes.fr