Fig. 2: Results of AsyT for training fully encapsulated DPNN.
From: Asymmetrical estimator for training encapsulated deep photonic neural networks
a The schematic of a fully encapsulated DPNN, where the analogue signal propagation is preserved for the entire deep network structure and only readout at the output layer. b Experiment setup for demonstrating the fully encapsulated DPNN. Two separate PIC devices are used to represent the deep neuron connectivity. The yellow arrows indicate the propagation of the optical signal. c Packaging of the PIC, showing the wire-bonded heaters and edge coupling to the fibre array. d A photo of the PIC device, with the enlarged components of the accumulator, splitter, and MZI cell. The schematic can be found in Supplementary Note 4. e The training result of the fully encapsulated DPNN. AsyT (dark blue line: AsyT train; orange line: AsyT test) shows significant improvement in performance compared to in-silico BP (faded blue line). f An estimation profile based on the theoretical estimation of the MZI unit response, showing the large mismatch between the estimation profile (light blue line) and the actual transmission behaviours of the MZI cells (faded grey lines in the background). PPM physical processing module.
