Fig. 1. (a) The geometry of multi-synaptic connections. (b) The experimental setup. The optical path is denoted by yellow, and the electrical path is black. The GDS file and chip micrograph of the DFB-SA neuron chip. (c1–c3) The optical spectrum of free-running DFB-SA, marked with different injection wavelengths. (d1–d3) The optical spectrum of DFB-SA with external optical injection corresponding to (c1–c3). (e1–e3) and (f1–f3) The input pattern and the corresponding output.

Fig. 2. The basic structure for supervised learning of photonic SNNs with temporal coding.

Fig. 3. (a) The process of digit pattern recognition based on a multi-synaptic photonic spiking neural network. (b) The output classification. (c) The basic principle for temporal coding SNN. (d) The STDP rule. (e) The mean distance of 4 different training methods with different N_syn.

Fig. 4. The output spike timing of all 10 output neurons (the row) at the input of all 10 patterns (the column) when N_syn=3.

Fig. 5. (a) The converged Distance with 1, 2, and 3 synapses, respectively. (b1–b2) Normalized weights for each of the two sets of synapses after convergence. (c1–c2) Normalized delays for each of the two sets of synapses after convergence. (d1) The converged Distance in the parameter space of weight mismatch and delay mismatch and (d2) projection of the contour with Distance=1.

Fig. 6. (a1–j1) The input spike patterns for all the 10 output neurons. (a2–j2) The correlated outputs of DFB-SA. (k1) The output spike interval for all 10 output neurons of 20 experiment trials. (k2) The output spike power for all 10 output neurons of 20 experiment trials.

Table 1. Accuracy of MNIST classification based on different models with different N_syn.