NIST Physical Measurement Laboratory
Tuesday, June 15, 2021, 3:00 EDT (1:00 MDT)
A video of this talk is available to NIST staff in the Math channel on NISTube, which is accessible from the NIST internal home page.
Abstract: In training ever larger deep neural networks, the communication of gradients of the synaptic weights between computational nodes in the data center is emerging as a bottleneck in energy and latency. Recent theoretical work has demonstrated that most of the important information in these matrices is contained in just a small part of the gradient’s eigenspectrum. Proposed methods of using low-rank approximations of the gradient suggest that this compression could save energy and time in network training but the known algorithms for doing so are essentially software based, and far slower than the hardware accelerated inference engines they are attempting to guide.
In this talk, I describe our efforts at building an efficient machine that can learn low-rank approximations of neural network gradients at roughly the same speed as the associated hardware-accelerated inference engine. Inspired at a basic level by neuron models where synapses compete for finite resources, the machine is comprised of interlocking, irregular trees of planar rotators, each associated with a unique rotation angle. These hardware-efficient rotators not only process the streaming neural network data but also manage a distributed representation of the local geometry in gradient space. We describe a simple algorithmic model of the machine, but also discuss issues introduced by the distributed non-local architecture and the problems that arise in the exploration of a non-Euclidean space. Finally, we discuss our results using models of this device to train a simple two-layer perceptron and contemplate future work, applications, and generalizations.
Bio: Matthew W. Daniels is a postdoctoral physicist in the Alternative Computing Group at the National Institute for Standards and Technology (NIST). He received a B.S. in Physics with a minor in Mathematics from Clemson University, and M.S. and Ph.D. degrees in Physics from Carnegie Mellon University. His dissertation work was on antiferromagnetic magnon physics. His postdoctoral work touches multiple topics in neuromorphic computing, with a focus on discovering and utilizing new information encodings that bridge the gap between traditional CMOS engineering and the energy efficient, native computing capabilities of magnetic and memristive nanotechnology.
Note: This talk will be recorded to provide access to NIST staff and associates who could not be present to the time of the seminar. The recording will be made available in the Math channel on NISTube, which is accessible only on the NIST internal network. This recording could be released to the public through a Freedom of Information Act (FOIA) request. Do not discuss or visually present any sensitive (CUI/PII/BII) material. Ensure that no inappropriate material or any minors are contained within the background of any recording. (To facilitate this, we request that cameras of attendees are muted except when asking questions.)