High performance and error resilient probabilistic inference system for machine learning

Choi, Jungwook

Permalink

https://hdl.handle.net/2142/88204 Copy

Description

Title

High performance and error resilient probabilistic inference system for machine learning

Author(s)

Choi, Jungwook

Issue Date

2015-07-17

Director of Research (if dissertation) or Advisor (if thesis)

Rutenbar, Rob A

Doctoral Committee Chair(s)

Rutenbar, Rob A

Committee Member(s)

• Chen, Deming
• Forsyth, David A.
• Shanbhag, Naresh R.

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Belief Propagation (BP)
• Field-Programmable Gate Array (FPGA)
• hybrid-core computing platform
• Markov Random Field (MRF)
• sequential tree-reweighted message passing (TRW-S)
• stereo matching
• Error resilience
• message passing
• algorithmic noise tolerance

Abstract

Many real-world machine learning applications can be considered as inferring the best label assignment of maximum a posteriori probability (MAP) problems. Since these MAP problems are NP-hard in general, they are often dealt with using approximate inference algorithms on Markov random field (MRF) such as belief propagation (BP). However, this approximate inference is still computationally demanding, and thus custom hardware accelerators have been attractive for high performance and energy efficiency. There are various custom hardware implementations that employ BP to achieve reasonable performance for the real-world applications such as stereo matching. Due to lack of convergence guarantees, however, BP often fails to provide the right answer, thus degrading performance of the hardware. Therefore, we consider sequential tree-reweighted message passing (TRW-S), which avoids many of these convergence problems with BP via sequential execution of its computations but challenges parallel implementation for high throughput. In this work, therefore, we propose a novel streaming hardware architecture that parallelizes the sequential computations of TRW-S. Experimental results on stereo matching benchmarks show promising performance of our hardware implementation compared to the software implementation as well as other BP-based custom hardware or GPU implementations. From this result, we further demonstrate video-rate speed and high quality stereo matching using a hybrid CPU+FPGA platform. We propose three frame-level optimization techniques to fully exploit computational resources of a hybrid CPU+FPGA platform and achieve significant speed-up. We first propose a message reuse scheme which is guided by simple scene change detection. This scheme allows a current inference to be made based on a determination of whether the current result is expected to be similar to the inference result of the previous frame. We also consider frame level parallelization to process multiple frames in parallel using multiple FPGAs available in the platform. This parallelized hardware procedure is further pipelined with data management in CPU to overlap the execution time of the two and thereby reduce the entire processing time of the stereo video sequence. From experimental results with the real-world stereo video sequences, we see video-rate speed of our stereo matching system for QVGA stereo videos. Next, we consider error resilience of the message passing hardware for energy efficient hardware implementation. Modern nanoscale CMOS process technologies suffer in reliability caused by process, temperature and voltage variations. Conventional approaches to deal with such unreliability (e.g., design for the worst-case scenario) are complex and inefficient in terms of hardware resources and energy consumption. As machine learning applications are inherently probabilistic and robust to errors, statistical error compensation (SEC) techniques can play a significant role in achieving robust and energy-efficient implementation. SEC embraces the statistical nature of errors and utilizes statistical and probabilistic techniques to build robust systems. Energy-efficiency is obtained by trading off the enhanced robustness with energy. In this work, we analyze the error resilience of our message passing inference hardware subject to the hardware errors (e.g. errors caused by timing violation in circuits) and explore application of a popular SEC technique, algorithmic noise tolerance (ANT), to this hardware. Analysis and simulations show that the TRW-S message passing hardware is tolerant to small magnitude arithmetic errors, but large magnitude errors cause significantly inaccurate inference results which need to be corrected using SEC. Experimental results show that the proposed ANT-based hardware can tolerate an error rate of 21.3%, with performance degradation of only 3.5 % with an energy savings of 39.7 %, compared to an error-free hardware. Lastly, we extend our TRW-S hardware toward a general purpose machine learning framework. We propose advanced streaming architecture with flexible choice of MRF setting to achieve 10-40x speedup across a variety of computer vision applications. Furthermore, we provide better theoretical understanding of error resiliency of TRW-S, and of the implication of ANT for TRW-S, under more general MRF setting, along with strong empirical support.

Graduation Semester

2015-8

Type of Resource

text

Permalink

http://hdl.handle.net/2142/88204

Copyright and License Information

Copyright 2015 Jungwook Choi

Owning Collections