Fig. 4: Neuromorphic computing simulation for classification and reconstruction using HOENN.

a LTP/LTD curves under varying linear polarization states (LP = 0°–90°). b Calculated nonlinearity of LTP (upper panel) and LTD (lower panel) as a function of polarization angle. c Schematic of the simulated optoelectronic network with 28 × 28 pixels input image from the Fashion-MNIST dataset, where the “T-shirt” category is utilized to train and test network performance. d Comparison of classification accuracy rates across multiple polarization angles (0°, 30°, 45°, 75°, and 90°) of the LTP/LTD pulses. The inset shows a representative input category mapping of the 784 synaptic weights to the output ‘T-shirts’ shown at the initial and final states of training. e The reconstruction results of the input image across training cycles (initial, 50, 250, and 500 epochs), with each cycle’s LTP/LTD pulses having different polarization angles (0°, 45°, and 90°). f Reconstruction accuracy as a function of training epochs for LP = 0°, 45°, and 90°. g Synaptic weight calculations, represented as the difference between device conductance (G⁺_n,m) and conductance normalizing factor (G^-_n,m), to illustrate the underlying learning dynamics, the inset shows the bidirectional weight update method