Fig. 4: 100-layer error-tolerant ONN.
a Deep ONNs were constructed using inter-chip computational propagation. High-speed inputs, organized into an 8-channel modulator array, were directed to the transmitter (inset) and loaded onto the single-layer chip. b Deep ONNs were trained for the complete ImageNet-1000 classification, with parameters including perturbed weights and perturbed nonlinearity downscaled to a 120-dimensional representation here. The grids in white/gray represent weights +1/−1, respectively. c Analysis was conducted for the deep ONNs involved in classifying the ImageNet-1000 dataset. The networks were tested against various levels of Gaussian noise featuring different standard deviations (Stds), focusing on both the ONNs with single-layer and multi-layer chip implementation as noise gradually distorted the layers. d Experiments with different depths and operational frequencies. The results indicate consistent improvements in performance with increased network depth for the ONN based on the single-layer chip, in stark contrast to the significant decline in classification accuracy seen in multi-layer ones at elevated frequencies. Error bars represent twice the Stds., with detailed values provided in [Supplementary Note 4 Computing capabilities of single-layer photonic chip]. A.u., arbitrary unit.
