Neuromorphic photonics

Neuromorphic photonics builds networks in which each element behaves like a neuron: it sums many weighted inputs and responds nonlinearly, sometimes with a spike, rather than executing instructions in a clocked pipeline. A photonic tensor core implements only the linear half of a neural network layer; neuromorphic designs are distinguished by putting the nonlinearity, and often the time dynamics, into the photonic or optoelectronic hardware too. The motivation is speed and latency. Optical and optoelectronic devices respond in picoseconds to nanoseconds, a million or more times faster than biological neurons, and a network that never leaves the physical domain avoids the digitization steps that dominate the cost of optical computing.

The most developed architecture is broadcast-and-weight. Each neuron's output is placed on its own wavelength and all are multiplexed onto a shared waveguide; at each receiving neuron, a bank of tunable microrings sets the weight of every wavelength, and a pair of photodiodes in balanced detection sums the weighted powers with sign. The photocurrent then drives a modulator whose nonlinear transfer curve serves as the activation function, and the modulator imposes the result on the neuron's own wavelength. This optical-electrical-optical neuron uses standard silicon photonic parts. All-optical alternatives take the nonlinearity from saturable absorption, gain saturation in a semiconductor optical amplifier, or the Kerr effect.

Spiking neurons use excitable lasers: devices such as lasers with a saturable-absorber section or VCSELs under optical injection, which rest quietly until a perturbation above threshold triggers a pulse of fixed shape, followed by a refractory period in which they cannot fire again. That is the qualitative behavior of a biological neuron, compressed to nanoseconds.

Reservoir computing avoids training most of the network. A fixed, randomly connected nonlinear system transforms the input, and only a linear readout is trained. In photonics a single nonlinear node in a delay loop can play the whole reservoir: the loop is divided into $N$ time slots of duration $\theta$, each acting as a virtual node, so the delay is $N\theta$. For 400 virtual nodes at 20 ps each the delay is 8 ns, 1.63 m of fiber at a group index of 1.468. Tasks demonstrated include equalizing distorted optical communication signals, where the input is already optical and low latency matters; that kind of front-end processing is the most plausible early use.

Obstacles remain between laboratory demonstrations and useful scale. A neuron's output must drive many others with enough power to spare, which requires gain in every layer; noise accumulates through cascaded analog stages; device variation means a network trained in software needs calibration or in-situ training on the hardware; and integrating thousands of tunable elements with their control electronics is unsolved at the density electronics already achieves.

References: B. J. Shastri et al., Nat. Photonics 15, 102 (2021); A. N. Tait et al., Sci. Rep. 7, 7430 (2017); L. Appeltant et al., Nat. Commun. 2, 468 (2011); P. R. Prucnal, B. J. Shastri, Neuromorphic Photonics (CRC Press, 2017).