Fig. 5. Experimental results of ACCEL for time-lapse tasks (video judgement).
a, Illustrations of traffic dataset with a vehicle moving in five different directions. We present one example from each of the five moving-direction categories, and one noised example in low-light conditions from the axial category. The original position, speed and size of the vehicle are set randomly. b, Experimental accuracies of ACCEL connected with single-layer digital NN, EAC connected with single-layer digital NN and OAC connected with single-layer digital NN with the incident light condition of 5 fJ μm−2 per frame (Supplementary Table 1). We use the sign function between EAC and OAC and the digital NN as the nonlinear activation. The pixel size of the phase mask in OAC is 9.2 µm, and the diffraction distance is 150 mm. c, Experimental classification accuracies of ACCEL connected with single-layer digital NN, EAC connected with single-layer digital NN and OAC connected with single-layer digital NN under different incident light powers. d, Experimental confusion matrixes of ACCEL connected with single-layer digital NN, EAC connected with single-layer digital NN and OAC connected with single-layer digital NN under low-light condition (0.08 fJ μm−2 per frame). The EAC and OAC are connected to a single-layer digital NN (48 × 5 neurons) after conversion with 1-bit comparators, the same as ACCEL for fair comparisons. a.u., arbitrary unit; Max., maximum; Min., minimum; Int., intensity.