Great read. The Conclusion and the Acknowledgements paragraphs are reproduced below. Well worth reading to understand the current driving forces behind spiking neural network technologies.
7 CONCLUSION
e task described here is an example of how neuromorphic hardware can have an impact on a much broader set of numerical applications than the community generally considers. Demonstratingthe ability for spiking neuromorphic systems to impact conventional numerical computing is important; by extending its impactbeyond cognitive applications we increase the likelihood of a longlasting eect on the computing eld. Notably, while we did notexert considerable eort on optimizing the results presented herefor either time or space, we already have evidence that neuromorphic hardware can be more ecient than conventional approacheswhen fully parallelizable Monte Carlo based are implemented. Forexample, with only minimal additional work and ignoring I/O considerations, we anticipate that we could potentially perform thecomplete simulations described above in parallel on a fully populated 32-Loihi chip Nahuku board in less than a minute.ICONS ’20, July 28–30, 2020, Chicago, IL, USA J. Smith, W. Severa, A. Hill, L. Reeder, B. Franke, R. Lehoucq, O. Parekh, and J. AimoneNot surprisingly, we observed several aspects of neuromorphichardware that will require further investigation. For one, the I/Ocosts of neuromorphic hardware will likely grow as a consideration.e costs of I/O are oen considered for streaming applicationssuch as real-time machine learning inference, but for the classof numerical simulations considered here, interactive I/O is notrequired, but tracking state—or in this case accumulating state—isrequired in order to properly evaluate the simulation. Since I/Owill likely continue to be a limiting factor, further processing ofsimulation outputs on the neuromorphic substrate is likely ideal.One way to do this is to implement the post-processing steps thatwe performed oine as neural circuits themselves and integratethem into a fully composed simulation [3].e second signicant consideration learned from the neuromorphic simulations is the potential impact of reduced precisionstochastic neurons on model performance. e stochastic stepsof our simulation are aected by the precision of internal neuronstates, precision of weights between neurons, and precision of therandom number generator. ese dierent components interactin complex, architecture-dependent ways and the implications ofthis reduced precision merit deeper exploration. At the same time,some of the benets of neural hardware—the ability to have morerandom number draws eectively in parallel—may be able to osetthese consideration.Nevertheless, these neuromorphic considerations should prove surmountable especially as future generation platforms become available. As non-anticipated applications such as these are explored, it will be increasingly evident what the potential implications of reduced precision and I/O are and whether the costs ofmitigation advocate for future hardware modications or improvedcircuit and algorithm design.
8 ACKNOWLEDGMENTS
is work was supported by Laboratory Directed Research andDevelopment program at Sandia National Laboratories. SandiaNational Laboratories is a multi-program laboratory managed andoperated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International,Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA-0003525. is paperdescribes objective technical results and analysis. Any subjectiveviews or opinions that might be expressed in the paper do not necessarily represent the views of the U.S. Department of Energy orthe United States Government. SAND2020-5296 O
BRN Price at posting:
12.0¢ Sentiment: Hold Disclosure: Held
(20min delay)