Source-linked AI summary
A silicon photonic modulator neuron
Alexander N. Tait, Thomas Ferreira de Lima, Mitchell A. Nahmias, Heidi B. Miller, Hsuan-Tung Peng, Bhavin J. Shastri, Paul R. Prucnal
TL;DR
Neuromorphic photonics lacked a silicon photonic neuron demonstrated to interact with like neurons, despite progress on isolated neurons and analog interconnects. The paper fabricates and demonstrates a modulator-class neuron in a conventional silicon photonic process, showing fan-in, nonlinear optical processing, configurable and inhibitory behavior, time-resolved processing, and autaptic cascadability. The device is presented as a component for fully integrated photonic neural networks on available silicon photonic platforms.
Problem
Silicon photonics had not yet produced a neuron demonstrated to network with like devices, although network-compatible neurons require fan-in, nonlinear processing, and cascadable outputs.
Method
The authors fabricate an optical-to-electrical-to-optical neuron comprising balanced photodetectors electrically connected to a microring modulator.
Results
The device demonstrates fan-in, high-gain optical-to-optical nonlinearity, indefinite cascadability, transfer-function configurability, inhibition, pulse compression, and time-resolved processing.
Takeaways & Limitations
Combined with microring weight banks, the modulator neuron supplies a component for broadcast-and-weight neural networks compatible with mainstream silicon photonic platforms.
Takeaways & Limitations
The autapse experiment uses a partially fiber-based feedback loop, introducing gain, frequency, latency, and robustness discrepancies relative to a fully integrated autapse.
Abstract
from arXiv · showhide
There has been a recently renewed interest in neuromorphic photonics, a field promising to access pivotal and unexplored regimes of machine intelligence. Progress has been made on isolated neurons and analog interconnects; nevertheless, this renewal has yet to produce a demonstration of a silicon photonic neuron capable of interacting with other like neurons. We report a modulator-class photonic neuron fabricated in a conventional silicon photonic process line. We demonstrate behaviors of transfer function configurability, fan-in, inhibition, time-resolved processing, and, crucially, autaptic cascadability -- a sufficient set of behaviors for a device to act as a neuron participating in a network of like neurons. The silicon photonic modulator neuron constitutes the final piece needed to make photonic neural networks fully integrated on currently available silicon photonic platforms.
I. INTRODUCTION
Neuromorphic photonics offers a route to high-bandwidth neural processing, but prior work had not demonstrated a silicon photonic neuron capable of networking with like devices. This work introduces an integrated modulator neuron with the capabilities required for network participation.
- Motivation: Neuromorphic electronics face limits for applications requiring nanosecond latency, motivating alternatives that move beyond purely electronic physics.The paper identifies offline accelerators, wideband RF pattern detectors, and fast feedback controllers as examples of such applications.
- Motivation: Photonic systems could operate 6–8 orders-of-magnitude faster than neuromorphic electronics while offering favorable multiplexing, dissipation, and cross-talk properties.
- Contribution: The demonstrated modulator neuron provides fan-in, high-gain optical-to-optical nonlinearity, and indefinite cascadability, with configurable transfer functions, inhibition, pulse compression, and time-resolved processing.The authors establish cascadability using an observable bifurcation in an autapse circuit.
- Research gap: Network-compatible photonic neurons must combine fan-in, nonlinear processing, and cascadable outputs that can drive multiple other neurons, including themselves.
- Contribution: The neuron integrates with microring weight networks in a broadcast-and-weight architecture, supporting independently configured connections on a silicon photonic platform.Each connection uses an independently configured microring weight, while wavelength-division-multiplexed carriers are detected without mutual interference.
II. METHODS
The neuron is an optical-to-electrical-to-optical circuit that subtracts two optical inputs, applies a microring modulator transfer function, and produces a remodulated output. Device characterization varies thermal bias, electrical bias, and optical inputs to probe configurable and inhibitory behavior.
- Device Description: The modulator neuron uses two balanced photodetectors electrically connected to a microring modulator to convert two optical inputs into a nonlinear optical output.The photodetectors subtract the inputs electronically, and the modulator remodulates a pump signal onto a new wavelength.
- Fabrication: The device contains silicon waveguides, an MRR with separate heater and high-speed modulator junctions, and a vertical P-I-N germanium photodiode.
- Device characterization: Thermal tuning shifts the resonance by 0.24nm/mW without significantly changing the response, whereas electrical bias changes peak depth and quality factor from 14.5k to 3.5k.Thermal tuning can therefore lock the MRR to a WDM channel while preserving the electro-optic response.
- Device characterization: Positive-port illumination adds photocurrent to the bias and blue-shifts the resonance, while negative-port illumination shunts bias current and enables complementary inhibitory effects.When both inputs are applied, the complementary effects can cancel, with red traces matching the no-light blue traces.
- Device characterization: The measured optical-to-optical gain is 2.16×10−2, increasing to an on-chip gain of 1.36 after removing 18.0dB fiber-to-chip insertion loss.Gain scales with pump power, which was held artificially low during the measurement.
- Device characterization: Injection modulation was used because depletion modulation was too weak in the non-optimized fabricated modulators, although injection modulation is slower.The paper reports bandwidths up to 6.25GHz for injection modulation and up to 40GHz for depletion modulators.
C. Experimental setup
The experiments use wavelength-division-multiplexed optical inputs, programmable delays, multiple waveform generators, and off-chip detection to characterize the neuron’s dynamic responses.
- Signal generation: Two input wavelengths, λ1=1546.4nm and λ2=1548.2nm, are wavelength-division multiplexed and power modulated to create distinct delayed input signals.A relative delay, ∆T, separates the two wavelength channels following the method of.
- Signal generation: The setup uses two analog signal generators and one binary generator to produce transfer-function, autapse, burst, and time-varying inputs.
- Output measurement: The neuron output is coupled off-chip, amplified and filtered through an EDFA-based signal-to-noise stage, then detected with a sampling oscilloscope.
- Experimental apparatus: The chip is aligned to a four-fiber grating-coupler array serving the IN+, IN–, PUMP, and OUT ports, with electrical biases supplied through probe tips and current sources.
III. RESULTS
The experiments characterize a silicon photonic neuron’s configurable nonlinear transfer functions and optical signal-processing behavior. The device produces several experimentally observed response shapes under different heater-bias conditions and supports time-resolved operation.
- The experiments examine nonlinear optical-to-optical conversion, transfer-function configurability, high-bandwidth operation, pulse compression, fan-in, inhibitory fan-in, and time-resolved processing.
- The experimental setup creates distinct delayed optical signals, routes the second wavelength to excitatory or inhibitory ports, and monitors the remodulated output.The balanced photodetectors generate complementary photocurrents that drive the MRR modulator together with bias current.
- A. Transfer functions: Changing the heater bias shifts the MRR resonance relative to the pump and changes the nonlinear response obtained from the photodetector–modulator pair.The response shapes correspond to different portions of the modulator’s nonlinear Lorentzian transfer function.
- A. Transfer functions: Six experimentally observed transfer-function shapes include sigmoid, ReLU, radial basis function, and quadratic responses under different heater-bias conditions.The input for the displayed transfer functions is a 40ns burst of a 100MHz carrier.
- The device also demonstrates reproduction, rectification, and pulse compression for time-varying optical inputs, extending its configurable nonlinear processing beyond slow transfer-function measurements.These behaviors are illustrated with a 25.0ns burst of a 1.0GHz RF carrier.
B. Response to high-bandwidth inputs
High-bandwidth experiments show that the neuron can transform fast RF inputs while preserving or compressing their temporal structure. The same device also supports linear and nonlinear excitatory or inhibitory two-channel operations.
- B. Response to high-bandwidth inputs: A 25.0ns pulse of a 1.0GHz RF carrier can be faithfully reproduced on a different optical carrier wavelength in the neuron’s linear regime.The reported voltage gain is 20%, although the authors caution that it is not necessarily representative of a fully integrated case.
- B. Response to high-bandwidth inputs: 20.4ns output bursts represent 19% compression when the ReLU is biased below its elbow.The thresholding behavior compresses the burst at the envelope level.
- B. Response to high-bandwidth inputs: 500ps full-width half-maximum input pulses become 324ps pulses, corresponding to 36% compression.The compression occurs in the ReLU response to the 1GHz pulse train.
- B. Response to high-bandwidth inputs: Pulse compression is attributed to the ReLU’s positive third-order nonlinearity, which places leading and lagging pulse edges below the ReLU elbow.The authors contrast this with more common optical saturation from negative third-order nonlinearities.
- B. Response to high-bandwidth inputs: Two-channel experiments demonstrate linear and nonlinear excitatory and inhibitory functions, including fan-in of two WDM signals onto one output wavelength.Inhibitory inputs counteract one another rather than merely inverting the excitatory case.
C. Response to multiple inputs
The modulator neuron performs fan-in across wavelengths, combines excitatory or inhibitory inputs, and applies configurable nonlinear transformations. With pulsed inputs, its response depends on coincidence timing, enabling enhancement, saturation, and inhibition.
- Multiple-input response: The neuron sums two wavelength-distinct inputs and converts them to a single output wavelength, demonstrating optical fan-in.The inputs can be combined through the same or complementary balanced-photodetector ports.
- Multiple-input response: Excitatory fan-in can be followed by nonlinear rectification, while inhibitory inputs counteract one another through complementary-port operation.The resulting optical signal can, in principle, drive other neurons, supporting multiple weighted inputs and outputs.
- Multiple-input response: The combined signal can implement envelope-detection operations proportional to (A + B)^2 and (A−B)^2, whose expansion contains a ±2AB correlation term.The paper relates these simple processing operations to RF signal-processing tasks such as dimensionality reduction and principal component analysis.
- Pulsed response: 157% of the linear solution: nonlinear enhancement occurs when coincident pulses cross the ReLU elbow, whereas saturation produces 56% of the linear solution.The enhancement experiment used 2 ns pulses; the normalized linear coincidence peak was 1.0.
- Pulsed response: Coincident inhibitory pulses produce approximately zero output, while non-coincident pulses generate complementary positive and negative perturbations.Inhibition is obtained by directing the second input to the IN– photodetector.
E. Cascadability
The study uses an autapse to test whether the modulator neuron can preserve signal integrity through feedback, providing evidence of indefinite cascadability. The experiment observes a cusp bifurcation as feedback gain changes, while highlighting differences between the fiber setup and a fully integrated circuit.
- Requirements and test: A cascadable neuron requires an operating point with large-signal gain of one, differential gain greater than one, and optical input and output at the same wavelength.An autapse exposes these conditions because its output is fed back as its input; differential gain above unity produces bistability.
- Experimental result: The observed transition between monostability and bistability corresponds to cascadability because autapse bistability occurs when differential gain exceeds unity.The experiment reports a cusp bifurcation for WF ∈[0, 1].
- Experimental setup: The autapse experiment multiplexes the neuron output with an external sawtooth input before feeding both into the positive photodetector.The feedback path includes fiber, polarization control, attenuation, an optical tap, and an EDFA.
- Experimental result: With blocked feedback, the neuron shows a regular input-output relation; with feedback, rising and falling inputs produce direction-dependent outputs.The feedback cases use WF = 1, whereas the baseline uses WF = 0.
- Why the autapse matters: An autapse provides stronger evidence of indefinite physical cascadability than a single feedforward link because identical upstream and downstream signals must drive the same neuron.The recurrent circuit is equivalent to a neuron driving an indefinite chain of identical neurons.
- Related constraints: Optical fan-in and cascadability face distinct challenges involving interference, phase dependence, loss, wavelength constraints, and electrical parasitics.The paper discusses mutually incoherent channels and O/E/O pathways as approaches to these constraints.
B. Non-spiking photonic neurons
The paper positions continuous-valued modulator neurons as a complementary alternative to predominantly spiking photonic neurons. It argues that their computational richness need not be substantially reduced, while noting practical bandwidth and fiber-feedback limitations in the demonstrated system.
- Motivation: Most photonic-neuron research has focused on laser-based spiking models, whereas modulator neurons are non-spiking, easier to fabricate, and lower-power on-chip.Spiking neurons may offer a theoretically richer processing repertoire.
- Computational scope: Continuous-valued neurons can represent mean firing rates directly with analog signals, offering an alternative to low-pass-filtered spiking representations.The paper identifies several situations motivating spiking neurons, including brain studies, amplitude-noise robustness, and temporal coding.
- Computational scope: The authors argue that continuous-valued modulator neurons do not sacrifice too much computational richness and may complement laser neurons across application domains.They also suggest that both neuron classes could coexist on one chip for different tasks.
- Limitations: The fiber autapse introduces gain, frequency, and robustness discrepancies relative to a fully integrated circuit, while fiber-to-chip loss weakens the direct gain-cascadability claim.The measured gain is 1.36 after accounting for fiber-to-chip insertion loss, and device calculations indicate g can exceed unity without that loss.
- Experimental interpretation: The kHz sawtooth frequency was selected to isolate equilibrium behavior from time-delayed and carrier-injection dynamics.Comparable responses across rising and falling cases support an equilibrium interpretation, while the fiber setup has limited environmental robustness.
- Relation to reservoirs: The fiber-based autapse differs from fiber-reservoir computing because it minimizes feedback dynamics rather than using them to generate complex dynamics.The two approaches therefore use similar hardware motifs for distinct computational purposes.
D. Further work
The paper identifies full integration, higher-bandwidth reproduction, and algorithm-to-hardware mapping as key next steps. These directions build on the demonstrated neuron and its compatibility with microring weight networks.
- Integration: A fully integrated autapse is a key future demonstration because it would avoid fiber-induced latency and fiber-to-chip coupling loss.The authors suggest that such an autapse could help quantify neuromorphic photonic energy consumption experimentally.
- Bandwidth: Higher-bandwidth experiments should address carrier-injection modulation and parasitic capacitance at the modulator connection.Suggested remedies include stronger depletion modulation, an on-chip resistor, and a series inductor.
- Algorithms and hardware: Existing neural algorithms and compilers could be combined with experimentally validated modulator-neuron transfer functions to map task specifications into hardware weights.The proposed workflow converts a high-level task into a weight matrix and maps those weights back to the physical network.
- Demonstrated foundation: The fabricated circuit directly demonstrated optical-to-optical nonlinearity, fan-in, and indefinite cascadability together in one integrated optoelectronic device.The circuit uses two photodetectors electrically connected to a microring modulator.
- System direction: Together with microring weight banks, the modulator neuron supplies the remaining component for broadcast-and-weight neural networks on a mainstream silicon photonic platform.The paper links this compatibility to feasibility, scalability, and economies of scale.
VI. APPENDIX: EQUIVALENCE OF CASCADABILITY AND BIFURCATION
The appendix derives when the photonic autapse is cascadable and shows that this condition is equivalent to crossing a bifurcation threshold observable in experiment.
- Cascadability requires the neuron's differential optical-to-optical gain, g, to exceed unity.
- The neuron combines a balanced photodetector with a microring modulator whose junction voltage controls modulator transmission.The modulator uses reverse-biased voltage-mode depletion modulation, with detector responsivity Rpd and biasing impedance Rb entering the circuit model.
- The modulator is biased at its maximum slope point to maximize modulation slope efficiency, with no-input operation at the bias voltage Vb.The equivalent π-voltage Vπ is defined for notational convenience through an equivalent Mach–Zehnder modulation slope.
- The experimental autapse is implemented with a microring modulator neuron and a reverse-biased modulator junction driven through impedance Rb.
- Combining device properties yields the pump power threshold required for cascadability.The text identifies this pump power as the minimum needed for cascadability.
B. Observation
The physical autapse is modeled as a one-dimensional dynamical system, and fixed-point stability is analyzed to identify the bifurcation corresponding to cascadability.
- The modulator voltage is the state variable of the physical autapse dynamical system.Photocurrent is induced by external or self-feedback optical input, with feedback satisfying Pin = Pout.
- The model's fixed-point bifurcation is found to be identical to the expression derived earlier in Eq. 10.
- Stability is assessed by Jacobian linearization rather than by solving the steady-state values directly.The steady-state solutions are not pursued because their parameter dependence is inelegant.
- In the one-dimensional system, a bifurcation occurs when the scalar Jacobian J equals zero and stability changes.Positive perturbations grow exponentially when J is positive, until saturation at another fixed point.
- The resulting transition is from monostability to bistability.
FUNDING
The work acknowledges support from the National Science Foundation.
- The research was supported by National Science Foundation grants ECCS 1247298 and DGE 1148900.