Source-linked AI summary
All-optical spiking neurosynaptic networks with self-learning capabilities
J. Feldmann, N. Youngblood, C. D. Wright, H. Bhaskaran, W. H. P. Pernice
TL;DR
Brain-like computing is constrained by separated memory and processing, motivating hardware neuromorphic alternatives. This paper demonstrates an all-optical spiking neurosynaptic network that learns and recognizes patterns, including four 15-pixel patterns and language detection above 90% accuracy.
Problem
Traditional processing separates memory and computation, limiting the speed and energy efficiency of brain-like computing.
Method
The paper integrates an all-optical spiking neuron with photonic synapses in a nanophotonic network for supervised and unsupervised learning.
Results
Above 90% accuracy was attained for language detection with 150 words, while the network successfully classified four 15-pixel patterns.
Takeaways & Limitations
The demonstrations support all-optical neurosynaptic networks as a route to optical-domain pattern recognition.
Takeaways & Limitations
The PCM cells can only be amorphized in a single step.
Abstract
from arXiv · showhide
Software-implementation, via neural networks, of brain-inspired computing approaches underlie many important modern-day computational tasks, from image processing to speech recognition, artificial intelligence and deep learning applications. Yet, differing from real neural tissue, traditional computing architectures physically separate the core computing functions of memory and processing, making fast, efficient and low-energy brain-like computing difficult to achieve. To overcome such limitations, an attractive and alternative goal is to design direct hardware mimics of brain neurons and synapses which, when connected in appropriate networks (or neuromorphic systems), process information in a way more fundamentally analogous to that of real brains. Here we present an all-optical approach to achieving such a goal. Specifically, we demonstrate an all-optical spiking neuron device and connect it, via an integrated photonics network, to photonic synapses to deliver a small-scale all-optical neurosynaptic system capable of supervised and unsupervised learning. Moreover, we exploit wavelength division multiplexing techniques to implement a scalable circuit architecture for photonic neural networks, successfully demonstrating pattern recognition directly in the optical domain using a photonic system comprising 140 elements. Such optical implementations of neurosynaptic networks promise access to the high speed and bandwidth inherent to optical systems, which would be very attractive for the direct processing of telecommunication and visual data in the optical domain.
Device fabrication
The nanophotonic circuits are fabricated by sequential electron-beam lithography, metallization, etching, and phase-change-material deposition on silicon nitride. GST is capped with ITO, crystallized thermally, and integrated into 1550-nm single-mode waveguides.
- Alignment markers: Electron-beam lithography uses a 100-kV system to define alignment-marker windows on a silicon wafer with 3300 nm oxide and 344 nm silicon nitride.The resist is developed in 1:3 MIBK:Isopropanol for 2 min before metal deposition.
- Alignment markers: A 5 nm chromium, 120 nm gold, and 5 nm chromium stack is evaporated by electron-beam PVD to form alignment markers.PMMA lift-off in acetone removes the resist and leaves the gold markers.
- Photonic structures: Photonic structures are patterned in maN 2403, reflowed at 105°C for two minutes, and transferred into silicon nitride by CHF3/O2 reactive-ion etching.Remaining resist is removed in oxygen plasma for 10 minutes.
- Phase-change material: GST is crystallized on a hot plate for about 10 minutes at 210°C, while the circuits use 1.2 µm-wide single-mode waveguides at 1550 nm.The waveguides define the operating optical platform for the fabricated photonic circuits.
Measurement setup
The measurement setup generated wavelength-multiplexed optical pulse patterns, coupled them into the on-chip neuron, and monitored transmission and output signals. Feedback paths enabled weight adjustment, while disabling the feedback amplifier placed the device in supervised learning mode.
- Optical read-out: Transmission read-out used a continuous-wave laser and four low-noise photodetectors D1–D4 monitored by a computer.The detectors were identified as New Focus Model 2011 units.
- Chip coupling: An aligned optical fiber array provided multi-port input through on-chip grating couplers, with polarization controllers used to optimize coupling efficiency.Pump and probe light counter-propagated through the device for beam-path separation.
- Pattern generation: Four cw lasers at different wavelengths were multiplexed, modulated by an EOM, electrically pulsed, amplified by an EDFA, demultiplexed, and guided to the chip.Arbitrary input patterns were selected by switching pump-light shutters.
- Feedback and learning: Output light was split between detector D0 and feedback ports F1–F4, with the feedback pulse used for weight adjustments.The output path used a cw laser, EOM, and EDFA before feedback distribution.
- Feedback and learning: Turning off the feedback amplifier put the device into supervised learning mode.The setup also included a four-neuron network configuration shown in supplementary material.