Source localization and separation
PhD thesis of Raphael frisch (2019)
Computers are without doubt one of the most important invention of the last century, whose impact cannot be overestimated. Over the years they became powerful, due to the constant optimization of their processors. With the growing need of computing power due to AI, processors have become faster than ever. However, since we are reaching the power wall, Moore’s law is coming to an end. Therefore, a young research community called rebooting computing is looking for alternative computation architectures.
In this work, we propose to use stochastic computing to build architectures dedicated to Bayesian inference aiming low-power consumption. We develop two machines, namely the Bayesian machine (BM) and the Bayesian sampling machine (BSM).
In this thesis, we look at two signal processing applications: Sound Source Localization (SSL) and Source Separation (SS). For SSL, we introduce three methods using the BM. The first one is working in the time-frequency domain and hence uses the Fourier transform. The second one is running entirely in the temporal domain. The third one is a multi-source localization approach based on the previous method.
We present a technique to speed up the stochastic computation by a factor of up to 103. Moreover, we designed an on-chip likelihoods computation mechanism to reduce the memory needs of our machine. Furthermore, we ran simulations and real world experiments to validate our methods. We made ASIC simulations to evaluate the power consumption. For the second problem, the source separation, we introduce a more general machine, the Bayesian sampling machine, which is based on the Gibbs sampling approach. We present a sampling method to solve source separation and run simulations to show the effectiveness of this technique.