Fig. 1. Conceptual illustrations of conventional photonic neurons, biological neural population, and the proposed comb-based photonic neural population (PNP). (a) Existing photonic neurons scale up in the spatial domain or temporal domain at the cost of physical complexity or response latency. The former is based on a single structure or devices, such as lasers [9–14" target="_self" style="display: inline;">–14], modulators [15], and resonators [16,17]. The latter relies on the temporal nodes, such as in reservoir computing [18] and optical pulses in an Ising machine [19]. (b) Schematic of the biological neural population with an input of motion direction. The neurons respond to different parts of the input range in a distributed manner (known as the bell-shape tuning curves). Complex input patterns can be represented by neural population coding and classified by an efficient readout process. (c) The proposed comb-based PNP. An injected optical frequency comb (OFC) can nonlinearly respond to input patterns in the frequency domain. Photonic tuning curves and comb power coding are exploited for processing complex input patterns. The comb powers can be obtained by a one-time readout, which ensures a low response latency.

Fig. 2. Experimental setup and corresponding schematic of the comb-based PNP. On the upper side, the input layer connects to the PNP layer by carrier coupling. The gain competition inside the PNP resembles the lateral inhibition connection in the biological counterpart [27]. The output layer represents the beat notes for the PNP activities monitoring. On the lower side, an electro-optical dual-OFC with slightly different comb spacings is generated by parallel PMs. The frequency-shifted OFC is optically injected into a DFB-LD and the PC is used for injection efficiency optimization. The output of the DFB-LD is monitored by an OSA and then detected in a low-speed PD with the other OFC. CW, continuous-wave laser; PM, phase modulator; AOM, acousto-optical modulator; PC, polarization controller; DFB-LD, distributed-feedback laser diode; OSA, optical spectrum analyzer; VOA, variable optical attenuator; PD, photodetector; ESA, electrical spectrum analyzer.

Fig. 3. Schematics and corresponding experimental results of the comb neuron response for a single input frequency. (a)–(d) The illustrative cases of no input, 6.6 GHz input, 9.6 GHz input, and 12.4 GHz input, respectively. The input frequency falls into response ranges of respective comb neurons and leads to a variation of the comb power. (e)–(h) Optical spectra of different input cases. The dotted box in red highlights the apparent decrease of corresponding comb power. (i)–(l) Electrical spectra of different input cases. The comb powers are monitored through the beat note powers. The center frequency and the spacing of the beat notes correspond to the frequency shift and frequency difference of the OFCs, respectively.

Fig. 4. Beat note powers dependent on input frequencies. The response range locates around the multiple times of the comb spacing. The inverse bell-shape response curve of each comb neuron resembles the tuning curve of biological neurons. The response width and amplitude are related to the initial comb power.

Fig. 5. Comb-based PNP activities of three input patterns with dual radio-frequency (RF) tones. The tones are randomly selected from three frequency bands. The PNP shows similar activities in response to the input pattern with approximate frequencies. Classification results can be consequently read out.

Fig. 6. Characterization of shapes of the photonic tuning curve. (a)–(c) Numerical results of the inverse bell-shape, dual-peak shape, distributed dual-peak shape, respectively. (d)–(f) Experimental results of the distributed dual-peak shape, dual-peak shape, and input-power dependence.

Fig. 7. Numerical results of classification of the complex input patterns. (a) Schematic of the classification task of four input patterns with 15 tones. (b)–(e) Optical spectra corresponding to the four input patterns, respectively. The sequence numbers of the lower comb powers are indicated.

Fig. 8. Experimental results of the pre-calibration method. (a), (b) The state comb power increases as the detuning frequency (fdetu) decreases. Hence, the comb power can be used as a sign of fdetu. (c) Time stability of the comb response corresponding to Fig. 4.