< BK21+ BEST Seminar Series Announcement> 

Time and Date : 11:00 ~ 12:00 Tuesday 06/25/2019

Place : D405, Engineering Building #4

Title : Accelerating Bayesian Inference for Structured Graphs Using Parallel Gibbs Sampling
Abstract:
Bayesian modeling and inference is an important class of machine learning, widely used for unsupervised or semi-supervised learning tasks. While successful deep learning models depend on large labeled data sets, Bayesian methods excel on problems with limited or no data for training and those that require representation and manipulation of uncertainties. One of the challenges when deploying Bayesian methods is the large amount of computation required to process high-dimensional integrals. As number of parameters increase, the inference on these models often become intractable. One of the ways to circumvent this is to use approximate inference techniques such as Markov chain Monte Carlo (MCMC) methods which is a class of algorithms for sampling from a distribution. In this work, we take Gibbs sampling, an MCMC method commonly used for Bayesian inference, and describe architectures for accelerators that allows parallelization of the algorithm on multiple samplers. We studied Gibbs sampling inference for Markov random fields (MRFs), which is a 2D-grid structured undirected probabilistic graphical model, commonly used in computer vision applications, such as image segmentation, image restoration, stereo matching and audio tasks like sound source separation. In this talk, we present FlexGibbs, a flexible architecture for Gibbs sampling inference on MRFs that allows us to generate an array of Gibbs samplers with desired number of labels per node for given hardware resources (i.e. FPGA size, available chip size, etc.) with a hardware scheduler capable of chromatic scheduling of the nodes in the graph. The chromatic scheduling guarantees that all the variables sampled in parallel are conditionally independent and converges to the target stationary distribution. Using a case study on real-time streaming sound source separation task, we show how the accelerator can be utilize to greatly improve speed and meet the stringent real-time latency requirement. Furthermore, by using asynchronous Gibbs sampling, a technique which allows asynchronous Gibbs updates between partitioned subgraphs, we are able to reduce the memory footprint for a more power and area constrained mobile System-on-Chip (SoC) setting with limited chip real estate. Using a TSMC 16nm FFC, we implemented PGMA, a programmable Gibbs sampling MRF accelerator on an ARM-based SoC that leverages two parallel Gibbs sampling techniques discussed above and show that it can run aforementioned computer vision tasks with image sizes up to 640x480 while achieving speedups of orders of magnitude over server-class CPUs while consuming tens of milliwatts.

Presenter: Gihyun Ko, Postdoctoral Fellow / Harvard University, USA

Host: Prof. Kim, Hanjun, Yonsei EEE