Date: Thursday 27 March 2014
Time: 16:00 - 17:30
Location / Room: Saal 1
Organisers:
Ian O'Connor, Lyon Institute of Nanotechnology, FR
Thomas Mikolajick, NamLab gGmbH, DE
Chair:
Michael Huebner, Ruhr Universitaet Bochum, DE
Co-Chair:
Ian O'Connor, Lyon Institute of Nanotechnology, FR
Challenges for acquiring and processing data from the real world includes the development of interfaces capable of extracting relevant information from massive sensor networks or from living organisms, sifting through the wealth of data to arrive systematically at a meaningful conclusion, and building hardware platforms suited to carry out these operations in an energy-efficient way. The first paper in this session looks at the necessarily complex processing of chemical information with hardware components that are capable of responding to various chemical conditions. Interfaces to living organisms are examined in the second paper, which discusses challenges and approaches for efficient detection of disease. In the third paper, novel hardware devices and architectures are explored for use in energy-efficient video analysis applications such as movement detection and face recognition. The fourth paper discusses handling of complex data with large-scale GPU-based recurrent networks, exploiting specific features of the data to improve energy efficiency.
| Time | Label | Presentation Title Authors |
|---|---|---|
| 16:00 | 12.1.1 | INTEGRATED CIRCUITS PROCESSING CHEMICAL INFORMATION: PROSPECTS AND CHALLENGES Speakers: Andreas Richter, Axel Voigt, René Schüffny, Stephan Henker and Marcus Völp, Technische Universität Dresden, DE Abstract The unbelievable properties of our information processing capabilities regarding the processing of big data, resilience, and energy efficiency are inspiration sources for the optimization and the rethinking of the principles of electronic information processing. Here, we present an approach of integrated circuits intended to solve chemical problems by active processing of chemical information. |
| 16:25 | 12.1.2 | INTERFACING TO LIVING CELLS Speaker: Rudy Lauwereins, IMEC, BE Abstract Recent advances in More than Moore technology enable close inspection of and even direct interfacing to living cells. This paper illustrates this through three use cases. In the first use case, the type or quality of billions of cells is quickly inspected in a fluidic medium. Secondly, the effect of potential drugs is monitored in neural cell cultures. In the third use case, neural brain activity is recorded in vivo using implantable electrodes to understand how the brain functions. |
| 16:45 | 12.1.3 | VIDEO ANALYTICS USING BEYOND CMOS DEVICES Speakers: Vijaykrishnan Narayanan1, Gert Cauwenberghs2, Donald Chiarulli3, Suman Datta4, Steve Levitan3 and Philip Wong5 1Penn State University, US; 2University of California at San Deigo, US; 3University of Pittsburgh, US; 4The Pennsylvania State University, US; 5Stanford University, US Abstract The human vision system understands and interprets complex scenes for a variety of visual tasks in real-time while consuming less than 20 Watts of power. The holistic design of artificial vision systems that will approach and eventually exceed the capabilities of human vision systems is a grand challenge. The design of such a system needs advances in multiple disciplines. This paper focuses on advances needed in the computational fabric and provides an overview of a new-genre of architectures inspired by advances in both the understanding of the visual cortex and the emergence of devices with new mechanisms for state computations. |
| 17:10 | 12.1.4 | ENERGY EFFICIENT NEURAL NETWORKS FOR BIG DATA ANALYTICS Speakers: Wang Yu, Boxun Li, Rong Luo, Yiran Chen, Ningyi Xu and Huazhong Yang, Tsinghua University, CN Abstract The world is experiencing a data revolution to discover knowledge in big data. Sequential data, such as the text, speech and video, are the primary sources of big data. The recurrent network is a powerful model to process sequential data because of the ability of capturing the long-term latent dependencies and features of the data. However, the difficulty of training a recurrent network, especially the huge requirement of computing power, makes the recurrent network fail to become a mainstream tool in mining big data. In this paper, we propose an efficient GPU implementation of large-scale recurrent network training. The proposed GPU implementation is based on a fast approximation technique of activation functions and a fine-grained two-stage pipeline architecture. We also propose a parallel realization of the stochastic gradient descent (SGD), one of the most popular but sequential algorithms for network training. The experiment results demonstrate that the proposed GPU implementation is able to realize at least 6x speedup on a signal GTX580 GPU compared with the CPU implementation on an Intel Xeon E5-2690 (16 cores) with MKL library. Meanwhile, the trained large-scale recurrent network can achieve the state-of-the-art performance on the Microsoft Research Sentence Completion Challenge, a challenge set for advancing language modeling. |
| 17:30 | End of session | |