SEPhIA shares one multi-wavelength laser across spiking neurons: a two-layer optoelectronic network reaches 91.35% on four-class image classification
Synopsis
HPE Labs proposed and validated SEPhIA, an optoelectronic spiking neural network architecture built on microring modulators and multi-wavelength lasers that lets many spiking neurons share one multi-wavelength light source, pushing the laser-per-neuron ratio below one. Prior photonic spiking work focused mostly on individual spiking devices or all-optical laser networks and rarely addressed optical power, crosstalk and footprint together at the architecture level. Using experimentally derived device parameters, a trainable optoelectronic spiking model reached 91.35% test accuracy on four-class spike-encoded image classification, about one percentage point behind a same-size fully connected spiking baseline at 93.65%.
Interpretation
Photonic spiking neural networks now have a whole-architecture design built around current photonic process constraints, rather than stopping at a single spiking device or an all-optical laser network. Earlier work in this direction mostly studied spiking devices themselves, networks of excitable lasers, or numerical models of large architectures, often leaving aside limited optical power, channel crosstalk and chip footprint. SEPhIA pairs microring modulators with multi-wavelength sources so that each spiking neuron needs fewer than one laser; at device level a Verilog-A time-domain co-simulation couples CMOS circuits to microrings, and at architecture level an optoelectronic spiking network is trained with experimentally measured device parameters.
The network is split into optical tiles following block-diagonal sparse weights, each tile sharing one multi-wavelength light source, with optical weighting on chip converted to electrical signals by balanced photodetectors. Each tile is driven by one multi-wavelength source, its input count set by the available wavelength-division-multiplexing channels and its output count adjustable; weighted optical signals are summed on paired balanced photodetectors to provide both excitation and inhibition. The demonstrated two-layer network has two parallel tiles in the first layer, each with a 16×16 microring weight bank and 16 spiking neurons, and a single 16×8 tile followed by 4 neurons in the second layer; training uses backpropagation over 35 discrete timesteps with a surrogate gradient for the non-differentiable spiking neurons.
Device parameters enter the trainable model directly, so a design-space study reports how classification accuracy shifts with wavelength-division-multiplexing spacing and microring quality factor and extinction ratio. Across nine parameter combinations, the widest 100 GHz channel spacing with high-Q, 15 dB extinction-ratio microrings gave the lowest validation loss, and validation loss generally rose as channel spacing narrowed. Each parameter set was trained in 5 repeated runs; at the densest 50 GHz spacing, 6 dB extinction-ratio microrings outperformed 15 dB ones, indicating complex trade-offs among device parameters under ultra-dense wavelength-division multiplexing.
Perspective
The demonstrated scope is simulation: device-level Verilog-A time-domain co-simulation of CMOS circuits coupled to microrings, and architecture-level classification by a two-layer, structurally sparse optoelectronic spiking network on four-class spike-encoded images, with device parameters taken from experimental data. The intended readers are researchers and engineers working on photonic neuromorphic hardware architecture and device co-design.
Classification accuracy comes from a simulation model that does not include thermal crosstalk among microrings and does not quantize resonance tuning in the weight bank; the authors list quantization, additional noise sources and different analog neural circuits as directions for future simulation.