Fig. 1. Photonic multi-synapse neural network. (a) Schematic of network architecture. (b) Photonic multi-synaptic connection implemented by input duplicates and multi-pathway diffractive propagation. (c) Experimental schematic.

Fig. 2. Experimental image classification. (a) Image examples from the datasets of MNIST, Fashion-MNIST, and CIFAR-10 (grayscale) in the phase-encoded form. (b) Schematic diagram of input duplicate array. (c) Test accuracy versus duplicate interval. (d) Reduction rate of classification error with respect to mono-synaptic connections. (e) Impact of ROI window on test accuracy. (f) Test accuracy under different camera exposure settings, with exposure time of 3 ms and exposure gains of 100% and 4000%. (g) Comparison of test accuracy between grayscale and color images considering joint training of RGB channels for CIFAR-10. (h) Test accuracies for the three datasets in duplicate array formats of 1×1, 5×5, and 9×9 with 10,000 hidden neurons. (i)–(k) Confusion matrices for the three datasets in the 9×9 format with 10,000 hidden neurons. (l)–(n) Test accuracy under different numbers of hidden neurons for the three datasets. CIFAR-10 (grayscale) is adapted in panels (c)–(f) and CIFAR-10 (RGB) is adapted in panels (h), (k), and (n).

Fig. 3. Test accuracy comparison. (a)–(c) Test accuracies on the three datasets for networks with mono-synaptic connections. (d) Test accuracies on the CIFAR-10 (grayscale images) dataset comparing the multi-synaptic connection effects.

Fig. 4. Experimental setup for photonic neural network.

Fig. 5. Neural network architectures. (a) General ELM neural network. (b) RNT-based neural network. (c) Optical model-based neural network. (d) Photonic neural network for grayscale image classification. (e) Photonic neural network for color image classification.