Photonic Spiking Actor Network

• Photonic Spiking Actor Networks are neuromorphic architectures that leverage optical spiking neurons and photonic matrix operations to achieve ultra-fast, energy-efficient decision making.
• They integrate device implementations like DFB-SA lasers, VCSELs, and CMOS-based circuits with programmable MZI mesh weight banks for scalable reinforcement learning.
• Experimental validations demonstrate sub-nanosecond inference and high energy efficiency, positioning these networks for real-time robotics, navigation, and network routing.

A Photonic Spiking Actor Network is a neuromorphic computing architecture that implements the policy network ("Actor") for reinforcement learning (RL) or other decision-making tasks using photonic hardware capable of spike-based computation. These networks combine ultrafast and highly parallelizable photonic matrix operations (linear transformations) with optical or optoelectronic spiking neuron elements to deliver high-throughput, low-latency, and energy-efficient inference in closed-loop control and decision-making problems. Architectures range from hybrids coupling photonic chips with electronic controllers to fully integrated large-scale photonic circuits, leveraging devices such as Mach–Zehnder interferometer (MZI) meshes, distributed feedback lasers with saturable absorbers (DFB-SA), vertical-cavity surface-emitting lasers (VCSELs), and passive microresonators (Lee et al., 2023, Xiang et al., 1 Feb 2026, Owen-Newns et al., 2022, Xiang et al., 2019, Xiang et al., 9 Aug 2025, Chen et al., 1 Feb 2026, Yu et al., 29 Nov 2025).

1. Photonic Spiking Neuron Models and Device Implementations

Photonic spiking neurons physically instantiate spiking dynamics that emulate, approximate, or extend biological neuron models, achieving orders of magnitude higher speed and potentially lower energy per operation compared to all-electronic implementations.

• DFB-SA lasers: A widely-adopted photonic spiking neuron leverages a distributed feedback laser with an integrated saturable absorber. The gain section, with carrier density $N_1$, and the saturable absorber (carrier density $N_2$) interact through a nonlinear dynamical system described by modified Yamada rate equations. When incident optical power (or electrical current) to the saturable absorber exceeds a programmable threshold, the device emits an ultrafast spike ($\sim$50–200 ps) and then enters a sub-nanosecond refractory regime (Chen et al., 1 Feb 2026, Xiang et al., 9 Aug 2025).
• VCSELs: Injection-locked single-mode VCSELs can exhibit excitable leaky integrate-and-fire behavior, governed by laser-carrier rate equations. Time-multiplexing in such devices allows the creation of virtual-node reservoirs suitable for rapid neuromorphic computation with >4–10 GHz node rates (Owen-Newns et al., 2022).
• Optoelectronic LIF neurons: CMOS-integrated circuits couple photodiodes and transistor-based subcircuits to realize leaky integrate-and-fire (LIF) neurons with adaptation, programmable via on-chip bias voltages that directly control parameters such as leak, threshold, adaptation strength, and refractory time constant. High programmability enables creating heterogeneous neuron populations akin to mixed-selectivity in biology (Lee et al., 2023).
• Passive photonic microresonators: All-optical spiking neurons can be constructed from passive microresonators side-coupled to waveguides, where nonlinearities (Kerr, thermo-optic, free-carrier) create excitability and cascadability in the absence of active (electrically pumped) components (Xiang et al., 2019).

A key feature across these platforms is the ability to support tunable or programmable neuron properties (e.g., membrane time constant $\tau_m$, firing threshold $V_{th}$, adaptation, refractoriness), essential for heterogeneity and functional expressivity in photonic SNNs.

2. Photonic Synaptic Weight Implementation and Network Architectures

Matrix-vector multiplication (MVM)—the core operation of neural networks—is mapped in photonic hardware through interferometric meshes, variable optical attenuators, or other integrated weight banks.

• MZI meshes: The standard approach for continuous, low-loss, and programmable weights is to use MZI meshes (e.g., Reck or Clements decomposition), where thermo-optic or electro-optic phase shifters encode real or complex-valued entries in the weight matrix $W$. Each node in a photonic mesh can perform linear transforms simultaneous across numerous channels. Networks can instantiate large hidden layers (e.g., 16×16 or 128×128) for sufficient policy expressivity (Xiang et al., 9 Aug 2025, Xiang et al., 1 Feb 2026, Yu et al., 29 Nov 2025).
• Weight programming: Devices use stochastic parallel gradient descent (SPGD) or similar in situ tuning: apply small random perturbations $\Delta\phi$ to phase shifters, measure the change in network fidelity to a target weight matrix, and update accordingly. This approach can be combined with lookup-table bias storage for individual neuron parameters ($V_{leak}$, $V_{th}$, etc.) (Lee et al., 2023, Xiang et al., 9 Aug 2025).
• Passive couplers/PCMs: For all-optical schemes, synaptic weights can be realized through tunable directional couplers, variable optical attenuators, or phase-change materials that serve as nonvolatile weights, supporting both feedforward and recurrent topologies (Xiang et al., 2019).

Network architectures include:

• Single-layer SNNs for low-dimensional tasks.
• Deep multi-layer actor–critic pipelines with photonic layers interleaved with spiking nonlinearities (Lee et al., 2023, Xiang et al., 9 Aug 2025, Xiang et al., 1 Feb 2026).
• Reservoir computing variants employing time-division multiplexed virtual nodes (Owen-Newns et al., 2022).

3. Reinforcement Learning Integration and Spiking Policy Training

Photonic Spiking Actor Networks are tightly integrated with modern RL algorithms, supporting policy sampling and gradient optimization with the physical constraints of photonically instantiated SNNs.

• Spiking PPO: Actor networks approximate $\pi_\theta(a|s)$ via spike-based coding, e.g., rate- or latency-coding output neurons. Policy updates employ proximal policy optimization (PPO) with a clipped surrogate objective, entropy regularization, and a value-function baseline. Gradients $\nabla_\theta L$ are estimated via surrogate-gradient backpropagation through spiking non-linearities, often using zero-inflated or fast sigmoid proxies for the threshold operation (Xiang et al., 1 Feb 2026, Xiang et al., 9 Aug 2025).
• TD3 and DDPG for continuous control: Continuous action spaces in robotic or navigation benchmarks are addressed through actor SNNs interfaced to conventional electronic (ANN) critic networks. The final nonlinear spiking activation is implemented in photonics (often a DFB-SA laser array), while the rest of the Actor–Critic loop is realized on CPU/FPGA (Chen et al., 1 Feb 2026, Yu et al., 29 Nov 2025).
• Hybrid and hardware-software co-design training: End-to-end training involves three stages: (1) software pre-training with surrogate backpropagation, (2) hardware mapping and in situ calibration of photonic linear layers via SPGD or fiducial measurement, (3) hardware-aware fine-tuning where linear photonic layer parameters are frozen during subsequent Actor optimization rounds. All methods maintain the standard RL data flow, but are adapted to match the operational semantics (non-negativity, device constraints) of photonic hardware (Xiang et al., 9 Aug 2025, Yu et al., 29 Nov 2025, Lee et al., 2023).

4. Experimental Demonstrations and Performance Metrics

Photonic Spiking Actor Networks have been experimentally validated on a range of RL and classification tasks, with performance metrics highlighting their advantages:

Platform/Task	Inference Latency	Energy per Operation	Task/Hardware Metrics
16×16 MZI+DFB-SA RL	$<320$ ps	$1.39$ TOPS/W (linear), $987.65$ GOPS/W (nonlinear)	CartPole reward 200, Pendulum reward $-250$, $\sim$98–100% accuracy (Xiang et al., 9 Aug 2025)
128×DFB-SA navigation	$191.2$ ps/inference	$0.78$ nJ/inference	Avg. reward 58.2, success 80%, error $<$0.06% (Chen et al., 1 Feb 2026)
MZI+SNN (continuous)	$120$ ps (photonic layer)	$1.39$ TOPS/W	HalfCheetah $5831$ reward ($23\%$ fewer steps) (Yu et al., 29 Nov 2025)
VCSEL-based RC	$<1$ ns per node	$\sim$10 mW total	$>97\%$ classification, $\sim10^{10}$ spike/s (Owen-Newns et al., 2022)
CMOS optoelectronic	$0.2$–$1$ ns per layer	$36.8$ fJ/spike (proj.), $1.18$ pJ/spike (meas.)	89.3% (Iris), sub-pJ per action (Lee et al., 2023)
Fat-tree SDN routing	$<50$ ns	$2$ pJ/decision	30–50% higher throughput vs. Dijkstra (Xiang et al., 1 Feb 2026)

These architectures demonstrate:

• Orders-of-magnitude speedup ($<1$ ns per inference) and energy reduction (femto- to picojoule per spike/inference), compared to digital electronic platforms.
• High inference accuracy ($>$98–100%) and reward equivalence to digital RL on standard benchmarks.
• Effective mapping to real-time robotic and networking applications.

5. Network Heterogeneity, Scalability, and Practical Considerations

Programmability and heterogeneity at the device and system level are vital for generalization and efficient learning.

• Neuron heterogeneity: By tuning per-neuron bias voltages (e.g., $V_{leak_m}$, $V_{th}$, $V_{leak_r}$, $V_p$ in optoelectronic circuits), it is possible to construct mixed-selectivity ensembles, improving network expressivity and task fit (Lee et al., 2023).
• Scalability: Photonic MZI meshes and DFB-SA arrays have been demonstrated up to $\sim$150 channels; future scalability relies on advances in photonic integration, SOI platforms, and multi-wavelength or multilayer photonics. Passive microresonator schemes offer massive parallelism at extremely low power, contingent on advances in on-chip tuning, thermal isolation, and nonvolatile weight banks (Xiang et al., 2019, Xiang et al., 9 Aug 2025, Chen et al., 1 Feb 2026).
• Integration with electronic control: Most state-of-the-art systems use hybrid architectures—photonic linear (weight) layers combined with electronic or optical nonlinearities, embedded in CPU/FPGA-accelerated RL pipelines for reward calculation and training. Full photonic learning loops are a prospect for future integration (Yu et al., 29 Nov 2025, Lee et al., 2023).

6. Limitations, Challenges, and Outlook

Photonic Spiking Actor Networks face several implementation challenges:

• Thermal management: Large-scale photonic circuits suffer from thermal crosstalk, requiring active stabilization (thermo-electric controllers, feedback) and advanced packaging to minimize drifts that impact coherent operations and weight retention (Xiang et al., 1 Feb 2026, Lee et al., 2023).
• On-chip learning: Realizing in situ learning with local photonic update mechanisms (e.g., phase-change materials, optomechanical actuators) is not yet fully mature. Most implementations require ex situ or hybrid weight adjustment protocols (Xiang et al., 2019, Lee et al., 2023).
• Device reproducibility and calibration: Process variations (e.g., resonance drift, phase-shifter nonidealities) mandate post-fabrication trimming, lookup-table calibration, and occasional compensation steps to match target transmission matrices (Yu et al., 29 Nov 2025, Lee et al., 2023).
• Spike encoding/decoding: Converting environmental or sensor data to spike trains at GHz rates, as well as decoding photonic output spikes for digital postprocessing, remains a practical throughput bottleneck, placing stringent requirements on ADC/DAC bandwidths and data movement (Xiang et al., 9 Aug 2025, Chen et al., 1 Feb 2026).

Despite these challenges, the convergence of scalable photonic components, programmable on-chip heterogeneity, and software–hardware co-design frameworks underpins the rapid progress of photonic spiking actor networks toward real-time inference and learning in robotics, network optimization, and autonomous navigation (Lee et al., 2023, Xiang et al., 9 Aug 2025, Xiang et al., 1 Feb 2026, Chen et al., 1 Feb 2026, Yu et al., 29 Nov 2025).