Proceedings Article | 13 March 2024
Bruno Garbin, Loredana Maria Massaro, Alexandre Bazin, Isabelle Sagnes, Konstantinos Pantzas, Sylvain Combrié, Alfredo De Rossi, Fabrice Raineri
Mainstream Machine learning (ML) leverages on a simplified model of the neuron, the Perceptron, which is efficiently implemented in software running on digital computers. Still, biological neurons process information by exchanging time-depending signals, e.g. spikes. Understanding how to harness ”neurons” closer in behaviour to their biological model, is fascinating but challenging. One of the challenges is related to scaling up the number of interconnected neurons. Photonics is regarded as a promising approach [1]. Particularly, the large bandwidth available in optical communication channels suggests applications in specialized computing tasks, where latency is critical. Aiming at an all-optical implementation, computing-related functions such as reconfigurable matrix multiplication [2] and nonlinear activation functions [3] are available in Silicon photonics. Semiconductor lasers have shown a neuron-like response such as excitability [4], i.e. the emission of a well-defined pulse as the excitation goes above a threshold, while microring-based ”neurons”, exploiting the thermo-optical nonlinearity, have been demonstrated in a silicon photonic circuit [5]. Here we consider a photonic crystal based semiconductor laser, heterogeneously integrated on top of a Silicon on Insulator (SOI) waveguide.
The nanolaser is composed by two sections for the gain and saturable absorber: a metal screen ensure selective pumping of the gain section with a CW laser beam. By a suitable choice of the parameters (Q-factor, gain vs
absorption ratio and pumping rate), the laser operates in different regimes: excitable, pulsing with variable rate (i.e. implements the Leak Integrate and Fire model of the neuron), bistable and CW. Excitability is shown in the figure: panels (c) to (e) relate to the response to a pulsed excitation as a function of its energy (here estimated before the input coupler to the silicon chip) when the pump is set just below the threshold of self-pulsing. Panel (f) shows the spiking probability, estimated as the fraction of traces where spikes are emitted (the event is detected when signal goes above -40 arb.u.). The spiking probability follows the expected trend with the excitation, already observed in VCSELs [6] and agrees very well with our stochastic implementation of the Yamada model of self-pulsing lasers.
This result paves the way to electrically pumped, interconnected spiking nanolasers, operating with manageable (0.1 mA) current levels and sub-nanosecond time scales.
References
[1] B. J. Shastri et al., “Photonics for artificial intelligence and neuromorphic computing,” Nat. Photon. 15, 102 (2021).
[2] Y. Shen et al., “Deep learning with coherent nanophotonic circuits,” Nat. Photon. 11, 441 (2017).
[3] A. Jha et al., “Reconfigurable all-optical nonlinear activation functions for neuromorphic photonics,” Opt. Lett. 45, 4819 (2020).
[4] H.J. W¨unsche et al., ”Excitability of a semiconductor laser by a two-mode homoclinic bifurcation” , Phys. Rev. Lett. 88, 023901 (2001).
[5] T. Van Vaerenbergh et al., “Cascadable excitability in microrings,” Opt. Express 20, 20292 (2012).
[6] F. Selmi et al., “Relative Refractory Period in an Excitable Semiconductor Laser,” Phys. Rev. Lett. 112 183902 (2014).
