Error-tolerant software adaptation of signal processing systems

Add to your list(s) Download to your calendar using vCal

• Yiannis Andreopoulos
• Thursday 18 October 2012, 16:00-17:00
• Room 503, Level 5, EEE Dept. .

If you have a question about this talk, please contact George A Constantinides.

Our research focuses on software frameworks that achieve increased processing throughput when producing approximate outputs. These can be used in mobile or high-performance computing systems when the precision of computation is not of critical importance (error-tolerant systems), when the input data is intrinsically noisy, or when the underlying hardware is not expected to be 100% reliable. Importantly, such software designs run on off-the-shelf computer hardware and they provide for on-demand reconfiguration depending on the input data and the precision and fault-generation profile of the underlying hardware (from full precision to noisy computation).

In our current work, this method has been developed for generic matrix multiplication (GEMM) and digital signal convolution, which are fundamental building blocks for several digital signal processing routines (factorization, system solving and template matching, to name a few). So far, our results have been very encouraging: for a state-of-the-art face recognition system, approximate computation behaves like a noise-reduction filter, which in fact improves the recognition accuracy and allows for up to two-fold increase in the processing throughput. This means that a high-performance computing cluster using the proposed GEMM design for face recognition can handle twice the number of input images per second in comparison to using a conventional state-of-the-art GEMM design (e.g. the GEMM call from Goto or ATLAS BLAS library). Other systems of this type can also take advantage of similar performance gains (a face recognition system is a particular case of an information retrieval/object identification system) and they are currently under investigation.

Perhaps more importantly, we are currently investigating the inherent error-tolerance capabilities of this new approach to establish new theoretical results, but also to build proof-of-concept error-tolerant designs and validate them in applications. An initial result of theoretical interest is the formation of the notion of computational capacity under a given distortion in the results of a high-performance linear algebra library based on the view of the underlying computer hardware as a noisy computational channel.

Related preprints: http://arxiv.org/abs/1201.3018, http://arxiv.org/abs/1110.5765

Brief Bio: Yiannis Andreopoulos obtained the Electrical Engineering Diploma and an MSc degree from the University of Patras, Patras, Greece. He obtained the PhD in Applied Sciences from the University of Brussels (Belgium) in May 2005. During his post-doctoral work at the University of California Los Angeles (US) he performed research on cross-layer optimization of wireless media systems, video streaming, and theoretical aspects of rate-distortion-complexity modeling for multimedia stream processing systems. From Oct. 2006-Dec. 2007, he was Lecturer at the Electronic Engineering Department of Queen Mary University of London. Since Dec. 2007, he is Lecturer at the Electronic and Electrical Engineering Department of UCL . During 2002-2004, Dr. Andreopoulos made several decisive contributions to the ISO /IEC JTC1 /SC29/WG11 (Moving Picture Experts Group – MPEG ) committee in the early exploration on scalable video coding.

Research interests: I am working in the fields of multimedia stream processing and coding, error-tolerant computing, signal processing & transform design, and wireless protocols for low-end systems (e.g. sensor networks).

This talk is part of the CAS Talks series.

This talk is included in these lists:

• Andrea Picciau's list
• CAS Talks
• COMMSP & CP list
• COMMSP Seminar
• Circuits and Systems Group: Internal Seminars
• Featured talks
• Room 503, Level 5, EEE Dept.
• SAP Talks

Note that ex-directory lists are not shown.