Autonomous agents require hardware, whether custom or commodity, to run complex neural models. In the past three years, I have pursued two complementary hardware approaches for hardware implementation of neural models at scale:
(1) in collaboration with my colleagues at Boston University and the industrial partner HP, I have developed and used software platforms exploiting mainstream (commodity) digital hardware ideal for fast algebraic implementations that can model the functions of certain brain processes, and
(2) in collaboration with Boston University partners, I have developed custom, low-power circuits to control learning and navigation in mobile robots.
Mobile, autonomous robots pose challenges not only in terms of intelligence, but also require low-power consumption and a small form factor to maximize battery life. Even virtual agents have constraints of processing or communication bottlenecks as well as power consumption, whether in a rack of processors or in the cloud. For some agents, instantiation in conventional hardware is appropriate, while others will require custom hardware.
My research leverages both approaches by adopting market-fueled progress in digital technology while staying current with the leading edge of novel neuromorphic computing devices. In both cases, principles from brain science will be embodied in silico. In the past two years, I have extended prior work in neural software to design, in collaboration with HP, a neural simulation computing platform to exploit digital processors (multicore, central processing units; CPUs, or graphic processing units; GPUs).
The resulting Cog Ex Machina, or Cog, platform has recently become a “Big Bet” at HP, namely one of the few projects that is believe to be transformational in the HP business model. Cog will leverage current technologies, in particular GPU clusters, as well as the imminent introduction of memristive memories and photonic interconnect, innovations that will reduce power consumption and make neural modeling more accessible. HP is using Cog to assist customers with challenging business problems that need smart algorithms and necessitate high-performance computing resources. The collaboration with HP is a currently funded project.
See also the code repository for code downloads.
- Massimiliano Versace, Heather Ames, Jasmin Leveille, Bret Fortenberry, and Anatoli Gorchetchnikov. KInNeSS: A modular framework for computational neuroscience. Neuroinformatics, 2008 Winter;6(4):291-309. PDF
- Snider G., Amerson R., Carter D., Abdalla H., Qureshi S., Leveille J., Versace M., Ames H., Patrick S., Chandler B., Gorchetchnikov A., and Mingolla E. (2011) Adaptive Computation with Memristive Memory. IEEE Computer 44(2), 21-28. PDF
Abstracts & Conference papers
- Raudies, F., Eldridge, S., Joshi, A., Versace, M. (2011). Reinforcement learning for visual navigation. NSF SLC PI meeting, Washington DC, November 2011.
- Eldridge, S., Joshi, A., Raudies, F., Versace, M. (2011). A Neuromorphic Hardware that Learns to Navigate Based on Optic Flow. Mark Motter NASA visit, Boston, MA March 31, 2011
- Samuel Kim and Vincent Kee (2011). Optical Flow Based Navigation. See link or PDF
- Kim, S., Kee, V., Joshi, A., Raudies, F. (2011). Optic flow based navigation. Boston University Research Internship in Science and Engineering Poster Session, August 12th, 2011, Boston, MA
- Kim, S., Kee, V., Joshi, A., Raudies, F. (2011). Optic flow based navigation using Gabor filters. Boston University Research Internship in Science and Engineering Poster Session, August 12th, 2011, Boston, MA
- Kim, S., Kee, V., Joshi, A., Raudies, F. (2011). Optic flow based navigation using correlation techniques. Boston University Research Internship in Science and Engineering Poster Session, August 12th, 2011, Boston, MA