Massachusetts Institute of Technology

Ntt Research, Inc.

Ryan Hamerly

Dirk Robert Englund

Saumil Bandyopadhyay

Cheng Peng

Christopher Louis Panuski

Zaijun Chen

The exponential growth in deep learning models is challenging existing computing hardware. Optical neural networks (ONNs) accelerate machine learning tasks with potentially ultrahigh bandwidth and nearly no loss in data movement. Scaling up ONNs involves improving scalability, energy efficiency, compute density, and inline nonlinearity. However, realizing all these criteria remains an unsolved challenge. Here, we demonstrate a three-dimensional spatial time-multiplexed ONN architecture based on dense arrays of microscale vertical cavity surface emitting lasers (VCSELs). The VCSELs, coherently injection-locked to a leader laser, operate at gigahertz data rates with a 7T-phase-shift voltage on the 10-millivolt level. Optical nonlinearity is incorporated into the ONN with no added energy cost using coherent detection of optical interference between VCSELs.

VCSEL-based coherent scalable deep learning

Programmable photonic circuits of reconfigurable interferometers can be used to implement arbitrary operations on optical modes, providing a flexible platform for accelerating tasks in quantum simulation, signal processing, and artificial intelligence. A major obstacle to scaling up these systems is static fabrication error, where small component errors within each device accrue to produce significant errors within the circuit computation. Mitigating errors usually involves numerical optimization dependent on real-time feedback from the circuit, which can greatly limit the scalability of the hardware. Here, we present a resource-efficient, deterministic approach to correcting circuit errors by locally correcting hardware errors within individual optical gates. We apply our approach to simulations of large-scale optical neural networks and infinite impulse response filters implemented in programmable photonics, finding that they remain resilient to component error well beyond modern day process tolerances. Our error correction process can be used to scale up programmable photonics within current fabrication processes.

Error correction for programmable photonics

Component errors prevent linear photonic circuits from being scaled to large sizes. These errors can be compensated by programming the components in an order corresponding to nulling operations on a target matrix X through Givens rotations X→TX, X→XT. Nulling is implemented on hardware through measurements with feedback, in a way that builds up the target matrix even in the presence of hardware errors. This programming works with unknown errors and without internal sources or detectors in the circuit. Modifying the photonic circuit architecture can reduce the effect of errors still further, in some cases even rendering the hardware asymptotically perfect in the large-size limit. These modifications include adding a third directional coupler or crossing after each Mach-Zehnder interferometer in the circuit and a photonic implementation of the generalized FFT fractal. The configured photonic circuit can be used for machine learning, quantum photonics, prototyping, optical switching/multicast networks, microwave photonics, or signal processing.

Self-configuration and error correction in linear photonic circuits

A reflective spatial light modulator (SLM) made of an electro-optic material in a one-sided Fabry-Perot resonator can provide phase and/or amplitude modulation with fine spatial resolution at speeds over a Gigahertz. The light is confined laterally within the electro-optic material/resonator layer stack with microlenses, index perturbations, or by patterning the layer stack into a two-dimensional (2D) array of vertically oriented micropillars. Alternatively, a photonic crystal guided mode resonator can vertically and laterally confine the resonant mode. In phase-only modulation mode, each SLM pixel can produce a &#x3c0; phase shift under a bias voltage below 10 V, while maintaining nearly constant reflection amplitude. This high-speed SLM can be used in a wide range of new applications, from fully tunable metasurfaces to optical computing accelerators, high-speed interconnects, true 2D phased array beam steering, beam forming, or quantum computing with cold atom arrays.

High-speed wavelength-scale spatial light modulators with two- dimensional tunable microcavity arrays

Deep learning performance is limited by computing power, which is in turn limited by energy consumption. Optics can make neural networks faster and more efficient, but current schemes suffer from limited connectivity and the large footprint of low-loss nanophotonic devices. Our optical neural network architecture addresses these problems using homodyne detection and optical data fan-out. It is scalable to large networks without sacrificing speed or consuming too much energy. It can perform inference and training and work with both fully connected and convolutional neural-network layers. In our architecture, each neural network layer operates on inputs and weights encoded onto optical pulse amplitudes. A homodyne detector computes the vector product of the inputs and weights. The nonlinear activation function is performed electronically on the output of this linear homodyne detection step. Optical modulators combine the outputs from the nonlinear activation function and encode them onto optical pulses input into the next layer.

Growing community of inventors

Cambridge, MA, United States of America

Ryan Hamerly