Ultrashort-Pulse Optical Neural Network Techniques for Brain-Scale Computing

Machine-learning technology enables web searching, social networks, e-commerce, and consumer products and is expected to become more and more prevalent in modern society. While incredibly effective, large neural networks are incompatible with traditional computers for which Moore’s law has hit physical limits, and memory accessing adds significant lag.

Specialized dense-data-optimized processors such as GPUs and TPUs now surpass these bottlenecks. However, current data centers based on these technologies are incredibly power hungry and scaling much larger presents a major challenge for the industry. Optics, which has already demonstrated superiority for interconnection tasks, is a potentially great fit for today’s incredible computing demands due to its low-loss propagation, high bandwidth, and lack of interference between neighboring channels.

While recent spatially extended optical neural network approaches have already demonstrated useful fast and efficient calculations for small neural networks, scaling size by several orders of magnitude to meet current demand appears to be an insurmountable technical challenge. Here we propose to leverage the low-loss and high-speed benefits of light to process information in time and space with massively time-multiplexed ultrashort optical pulses.

In this way >billion-scale model sizes can be processed in real-time with few spatial elements, bypassing the limits of current optical computing architectures with slow reconfigurability and poor spatial scaling. Beyond technological impact, this project will train two PhD students in an area of large technological importance at the interface of machine learning, ultrafast nonlinear optics and advanced optical technologies, and the PIs will integrate this important platform into the regular curriculum as well as into the extra-curricular optics summer-school program at the University of Rochester.

Technical description

Optics is being recognized as a potential solution for today’s incredible artificial neural network (ANN) computing demands due to its low-loss propagation, high speeds, and lack of interference, enabling parallel, efficient and fast information processing. In the last few years, several approaches to optical ANNs have been demonstrated. ANNs require linear matrix vector multiplications as well as nonlinear thresholding of the products.

Optical thresholding is achieved through modulators, lasers, detectors and several other techniques and the linear weight matrix vector calculations are processed commonly using micro-ring resonator arrays, Mach Zehnder interferometer arrays, and diffractive optics networks. These systems are challenging, however, because they are complex, highly calibrated nano-photonic devices with a limited number of input vector elements.

The research proposed here leverages the high speed, high bandwidth, and low loss properties of light by multiplexing information in the time domain, in addition to the spatial domain. In this way >billion-scale model sizes can be processed in real-time with few spatial elements, bypassing the limits of current optical computing architectures with slow reconfigurability and poor spatial scaling.

Through this approach, vector values coded in the amplitude of the electric field of a pulse are effectively processed with a novel matrix scaling algorithm, two in-memory all-optical accumulators, and a single nonlinear activation unit, without requiring intermediate electronics that limit the speed and efficiency of other time domain platforms. Finally, this system has unique advantages for extremely sparse matrices unavailable with prior techniques.

Beyond technological impact, this project will train two PhD students in an area of large technological importance at the interface of machine learning, ultrafast nonlinear optics and advanced optical technologies, and the PIs will integrate this important platform into the regular curriculum as well as into the extra-curricular optics summer-school program at the University of Rochester.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.