Extended Data Fig. 3. Evaluation of the robustness of adaptive training obtained by numerical simulations.
a, The testing accuracy of ACCEL under the condition of fabrication errors in the OAC phase mask with adaptive training of different scales of training datasets. The phase pattern in OAC is disturbed by Gaussian noises with a mean value of zero and standard deviation of 0.26π to simulate the fabrication error. b, The testing accuracy of ACCEL under the condition of lateral misalignment between OAC and EAC with adaptive training of different scales of training datasets. The OAC and EAC are misaligned by shifting one column horizontally. c, The testing accuracy of ACCEL under the condition of rotation misalignment between OAC and EAC with adaptive training of different scales of training datasets. The OAC and EAC are misaligned by rotating clockwise by 5 degrees around the centre. All these results are tested on MNIST dataset. The scales of training dataset are 100, 500, 1,000, 1,500, 2,000, 4,000, 7,000, 10,000, 20,000, 40,000 and 60,000 images. To match the parameters in experiments, we set the pixel size of the phase mask in OAC as 9.2 µm and the diffraction distance as 150 mm here, and the neuron number in OAC is 264 × 264.