Source-linked AI summary
Deep Learning with Coherent Nanophotonic Circuits
Yichen Shen, Nicholas C. Harris, Scott Skirlo, Mihika Prabhu, Tom Baehr-Jones, Michael Hochberg, Xin Sun, Shijie Zhao, Hugo Larochelle, Dirk Englund, Marin Soljacic
TL;DR
Conventional von Neumann hardware is inefficient for neural-network computation, motivating faster and more energy-efficient architectures. This paper proposes a coherent optical neural network using programmable nanophotonic circuits for matrix multiplication and nonlinear activation. The architecture is experimentally demonstrated on vowel recognition and is reported to offer substantially faster forward propagation with energy-efficient scaling, while hardware noise and data movement remain important boundaries.
Problem
Von Neumann computers are inefficient for neural-network workloads, creating a need for faster and more energy-efficient forward propagation.
Method
The paper implements neural-network matrix multiplication and nonlinear activation with coherent optical circuits on a programmable nanophotonic processor.
Results
At least two orders of magnitude faster forward propagation is reported under certain conditions, with vowel recognition reaching 138/180 cases (76.7%) experimentally.
Takeaways & Limitations
Optical forward propagation can provide high-speed, energy-efficient neural-network computation, with future phase-change materials potentially eliminating power needed to maintain phase settings.
Takeaways & Limitations
Performance is constrained by analog hardware errors including finite precision, photodetection noise, and thermal cross-talk, while data-movement energy remains unaddressed.
Abstract
from arXiv · showhide
Artificial Neural Networks are computational network models inspired by signal processing in the brain. These models have dramatically improved the performance of many learning tasks, including speech and object recognition. However, today's computing hardware is inefficient at implementing neural networks, in large part because much of it was designed for von Neumann computing schemes. Significant effort has been made to develop electronic architectures tuned to implement artificial neural networks that improve upon both computational speed and energy efficiency. Here, we propose a new architecture for a fully-optical neural network that, using unique advantages of optics, promises a computational speed enhancement of at least two orders of magnitude over the state-of-the-art and three orders of magnitude in power efficiency for conventional learning tasks. We experimentally demonstrate essential parts of our architecture using a programmable nanophotonic processor.
OPTICAL NEURAL NETWORK DEVICE ARCHITECTURE
The optical neural network maps conventional neural-network layers onto optical interference and nonlinearity units, using photonic components for matrix multiplication and activation. Singular-value decomposition enables the interference unit to implement arbitrary real-valued matrices, while optical nonlinearities provide activation functions.
- Each neural-network layer performs a linear combination followed by a nonlinear activation function.
- Optical pulses encode signals in amplitude and pass through an optical interference unit followed by an optical nonlinearity unit.
- Singular-value decomposition decomposes a general real-valued matrix into UΣV∗, with unitary factors and a non-negative diagonal matrix.
- Optical beamsplitters and phase shifters implement the unitary transformations, while attenuators or optical amplification materials implement Σ.
- Optical nonlinearities such as saturable absorption and bistability implement activation through the relation Iout = f(Iin).
EXPERIMENT
The experiment programmed a coherent nanophotonic processor to implement a multilayer optical neural network for vowel recognition. The hardware achieved 76.7% test accuracy, while practical analog nonidealities and an electronically implemented nonlinearity constrained the proof-of-concept demonstration.
- 360 four-feature vowel datapoints from 90 speakers and four phonemes were split evenly between training and testing.
- The processor comprised 56 Mach–Zehnder interferometers and 213 phase-shifting elements fabricated on a silicon-on-insulator photonics platform.
- SVD-based arbitrary matrix multiplication was realized with an SU(4) core and a non-unitary diagonal matrix multiplication core.
- Finite phase-setting precision, photodetection noise, and thermal cross-talk limited the optical computation’s effective precision.
- In the proof-of-concept, the nonlinear transformation was implemented electronically by detecting optical intensities and injecting outputs into the next interference stage.
- 138/180 cases (76.7%) were correctly identified by the programmed ONN, compared with 165/180 (91.7%) in simulation.
DISCUSSION
The discussion frames fast, low-power forward propagation as a central challenge and considers how optical hardware could address it. It also examines noise, scaling, and potential extensions of the architecture.
- High-speed, low-power forward propagation is central because data-center resources are heavily spent on test-time prediction and slow inference limits applications such as self-driving cars.
- At 100 GHz detection, an ONN with m layers of N×N matrix multiplication performs R = 2m · N2 · 1011 operations/s.The expression assumes N nodes, m layers, and a typical 100 GHz photodetection rate.
- Photodetection and phase-encoding noise are identified as dominant error sources in the analog ONN implementation.The paper evaluates correctness under varying phase-encoding error σΦ and photodetection error σD.
- The noise analysis uses simulated and experimental vowel-recognition results, including error-free training matrices and correctness contours across σΦ and σD.
- Current photonic technologies can realize ONNs exceeding 1000 neurons, with demonstrated circuits containing up to 4096 optical components and possible further 3-D scaling.Processing inputs through multiple patches over time is described as another route to larger effective networks.
CONCLUSION
The proposed architecture is presented as applicable beyond the demonstrated network to algorithms that rely heavily on matrix multiplication and nonlinear activation. Realizing its full advantage still requires addressing data-movement energy through further optical integration.
- The architecture could be applied to convolutional and recurrent neural networks because they heavily use matrix multiplications and nonlinear activations.
- Data movement remains an outstanding source of power dissipation, so integrated optical interconnects and computing units need further exploration.
METHODS
The methods evaluate ONN fidelity and noise robustness by perturbing trained unitary matrices and simulating noisy forward propagation. Correctness distributions are obtained across phase-encoding and photodetection noise levels.
- The study evaluates the SU(4) core using a fidelity metric based on experimental and simulated normalized optical-intensity distributions.
- The noise simulation varies phase-encoding noise σΦ and photodetection noise σD to assess the performance of trained matrices.
- Each trained 4 × 4 unitary matrix is represented by phase parameters, which are then perturbed before re-encoding the matrices.
- Forward propagation is performed on test data using the perturbed matrices, with random photodetection errors added after each matrix multiplication.
- Correctness distributions for each pair of phase-encoding and photodetection noise levels are obtained by repeating the simulation 50 times.