Thermodynamic Computing Workshop – Call for White Papers

The Computing Community Consortium (CCC) will hold a workshop from January 3rd to 5th, 2019 in Hawaii to create a vision for thermodynamic computing, a statement of research needs, and a summary of the current state of understanding of this new area. Workshop attendance will be by invitation only and travel expenses will be available for select participants. We seek short white papers to help create the agenda for the workshop and select attendees.

Thermodynamics has a long history in the engineering of computing systems due to its role in power consumption, scaling, and device performance [1],[2]. In a different context, thermodynamically motivated algorithmic techniques are prevalent and highly successful in areas such as machine learning [3], simulated annealing [4], and neuromorphic systems. The foundational thinking underlying much of the existing technology derives largely from equilibrium properties of closed thermodynamic systems. We aim to foster a community to extend these foundations into the domain of non-equilibrium thermodynamics toward the development of a new class of technologies that we call open thermodynamic computers.

The overall intuition is that striving for thermodynamic efficiency is not only highly desirable in hardware components, but may also be used as an embedded capability in the creation of algorithms: can dissipated heat be used to trigger adaptation/restructuring of (parts of) the functioning hardware, thus allowing hardware to evolve increasingly efficient computing strategies? Recent theoretical developments in non-equilibrium thermodynamics suggest that thermodynamics drives the organization of open systems as a natural response to external input potentials; that is, that these systems adapt as they dissipate energy, enter low dissipation homeostatic states and as a result ‘learn’ to ‘predict’ future inputs [5],[6]. For example, lower bounds on thermodynamic efficiency in driven systems (away from equilibrium), indicate that systems have to remain relevant, predictive information in order to be thermodynamically efficient [7],[8]. This strategy is, of course, the same as what is followed in machine learning (and, in general, in science): predictive inference [9]. This interesting connection between energy efficiency and information processing inspires us to bring together researchers in the various disciplines with the goal of building the foundations that would allow us to build radically different computing systems.

This CCC workshop will gather a set of leading researchers working to define open thermodynamic computers, describe the reasons that they should be studied, enumerate the major challenges that lay before us, and create a strategy for a way forward. We seek a diverse group of physical theorists, electrical and computer engineers, and electronic / ionic device researchers with a strong understanding of thermodynamics.

For this workshop, we request whitepapers of no more than two pages. Topics of interest include, but are not limited to:

• Non-equilibrium thermodynamic theory and its implications for self-organizing computing technologies, such as fluctuation-dissipation theorems and their ramifications regarding homeostasis, learning by prediction, stability, and plasticity.
• Novel device and component concepts that change and retain state (learn) in response to the thermodynamics of their environment, which may be suitable as the evolvable elements for future for thermodynamic computers.
• System concepts that integrate human-directed programmability and thermodynamically evolvable elements.
• Concepts to integrate and embed thermodynamic computers into an open environment of electrical and information potential.
• Non-equilibrium, thermodynamically inspired machine learning techniques and algorithms, such as simulated annealing in complex non-equilibrium environments and unsupervised learning by prediction.
• Challenge problems to motivate the development of thermodynamic computers

Topics out of scope include:

• Machine learning accelerators.
• Established machine-learning techniques.
• Brain simulations / models / simulators.
• Novel, programmable computing systems.

Authors of white papers may be asked to participate in teleconferences to develop an agenda prior to the workshop.

Please submit your white paper by October 5, 2018 to this registration form. For more information, please visit the workshop website. The organizing committee will notify the selected attendees by October 26th. Should you have any questions, please contact Khari Douglas at kdouglas@cra.org.

Thank you,
The Organizing Committee:
Tom Conte, Prof. of Computer Science, Georgia Inst. of Technology
Erik DeBenedictis, Sandia National Laboratories
Natesh Ganesh, Univ. of Massachusetts, Amherst
Todd Hylton, Prof of Practice ECE, UC San Diego
Susanne Still, Prof. of Information & Computer Science, Univ. of Hawaii, Minoa
John Paul Strachan, Hewlett Packard Labs, HPE
Natesh Ganesh, Univ. of Massachusetts, Amherst
Stan Williams

References:

1. Frank, Michael P. “The physical limits of computing.” Computing in Science & Engineering 4.3 (2002): 16-26.
2. Lloyd, Seth. “Ultimate physical limits to computation.” Nature 406.6799 (2000): 1047.
3. Dayan, Peter, et al. “The helmholtz machine.” Neural computation 7.5 (1995): 889-904.
4. Černý, Vladimír. “Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm.” Journal of optimization theory and applications 45.1 (1985): 41-51.
5. N. Ganesh, “A Thermodynamic Treatment of Intelligent Systems,” 2017 IEEE International Conference on Rebooting Computing (ICRC), Washington, DC, 2017, pp. 1-4.
6. Perunov, Nikolay, Robert A. Marsland, and Jeremy L. England. “Statistical physics of adaptation.” Physical Review X6.2 (2016): 021036.
7. Still, Susanne, et al. “Thermodynamics of prediction.” Physical review letters 109.12 (2012): 120604.
8. Still, Susanne. “Thermodynamic cost and benefit of data representations.” arXiv preprint arXiv:1705.00612 (2017).
9. Still, Susanne. “Information bottleneck approach to predictive inference.” Entropy 16.2 (2014): 968-989.