Special Sessions

➠ SUBMIT YOUR MANUSCRIPT HERE

Transprecision Architectures and Memories

Session Organizers: Dionysios Diamantopoulos (IBM) and Christian Weis (Technische Universität Kaiserslautern)

Abstract: Approximate computing has been recognized as an effective set of techniques to overcome the energy scaling barrier of computer systems. Such techniques rely on different layers of the computing stack to exploit the intrinsic error resilience of algorithms in many application domains such as signal processing, multimedia, data analytics and machine learning. Indeed, fully accurate arithmetic in specific phases of a computation in those applications may have only a marginal effect on output quality, especially if combined with system-level design. Thus, accurate execution may be traded off for lower energy consumption by providing the ability to scale supply voltage below nominal values or to use lower precision arithmetic (e.g. 16, 8 or 1-bit). Inspired by approximate computing, recently, transprecision computing has emerged as a computing paradigm that combines all the layers from the application to the specifically tuned hardware that supports a variety of precision settings, respective computations and configurations to access the transprecision memories. This special session intends to cover the emerging computing subsystem and memory subsytem in an inter-disciplinary effort to bring together researchers to discuss challenges, risks and opportunities of transprecision computing with focus on architectures, modeling and simulation.

◊◊ ◊◊ ◊◊

Machine learning implementations

Session Organizer: Wonyong Sung (Seoul National University, KR)

Abstract: Deep neural networks have shown great successes across a wide variety of tasks. However, the computational demand of deep learning is usually very high compared to conventional machine learning algorithms. GPGPUs (General Purpose Graphics Processing Units) have been most widely used, however more efficient solutions are also emerging in these days. This special session is intended to cover the emerging hardware and software solutions for high performance and efficient deep neural network training and also inference. Especially, the following topics are of great interests: high performance programmable deep neural network training and inference systems, custom hardware and FPGA based architecture, low-complexity neural network design such as quantization and pruning, and software solutions.

◊◊ ◊◊ ◊◊

Insights from Negative Results

Session Organizer: Karol Desnos (IETR, France) and Shuvra Bhattacharyya (University of Maryland, College Park , USA & IETR, France)

Motivation: Frequently, researchers choose to develop an idea because early or theoretical results seem promising. When scaling up, or when implementing the idea in a realistic scenario, the results are not as good as expected. Unfortunately, publishing such results is very difficult, because the reviewing process generally promotes positive results. When such a scenario occurs, scientists sometimes try to publish the idea anyway, by "disguising" the results in a positive way: by deliberately overlooking negative aspects of their results or by choosing an unusual evaluation metric.

Session objective:: The special session is dedicated to the presentation of ideas that lead to negative results. Contributions are expected to present the original idea, to explain why this idea was expected to provide good results, to present how and why it failed, and to discuss the presence (or absence) of solutions to solve encountered issues.


The objectives of this special session are fourfold:

  • To give credit to the research process that lead to the negative result.

  • To avoid disguising a negative result into a good one, which may trick other researchers into implementing an idea only to realize its actual negative side-effects later.

  • To help prevent other researchers from trying to develop this idea. If the idea seemed good at first, other researcher may have it one day. By publishing the idea and its unexpected negative results, you may save some time for other researchers.

  • To promote good ideas behind the negative results. Some ideas producing negative results may need some adapting or fine tuning before they can achieve their potential. A negative results forum provides a way to disseminate these kinds of ideas so they don't just get buried.

Previous edition:: This is the second edition of a SAMOS special session devoted to interesting negative results. The first edition was part of SAMOS 2018. As examples of previous papers under this theme, two selected papers from the first edition can be found at the following links:

◊◊ ◊◊ ◊◊

European Projects

Session Organizer: Dimitrios Soudris (NTUA, GR)

Abstract: This special session welcomes presentations on the latest results of European projects. Projects from all aspects of the Horizon 2020 Work Programme are invited to submit manuscripts through the SAMOS submission portal. The special session will especially put forward Future and Emerging Technologies, Leadership in Enabling and Industrial Technologies, as well as key proposals on European Societal Challenges.

◊◊ ◊◊ ◊◊

 

To submit a paper to either one of the conference tracks or the special session(s), select the proper option during the paper submission process. You can submit your paper HERE (Softconf - Open).