Fig. 3: Performance and analysis of AsyT with Iris dataset.
From: Asymmetrical estimator for training encapsulated deep photonic neural networks
a The training results for AsyT (dark blue line: train; orange line: test), pseudo-IP-BP (red line) and in-silico BP (faded blue line). AsyT retrieves performance near the maximum ideal BP for both training and testing. On the other hand, the performance of in-silico BP and pseudo-IP-BP shows degradation. b The confusion matrix for the testing data of the AsyT method. c AsyT quickly reduces the alignment angle difference at the early stage of training and maintains the value for subsequent epochs. d Repeated experiments with three additional estimation profiles, showcasing that the AsyT method can generalized to different estimation profiles. All three estimation profiles achieve a training accuracy of 96%.
