Fig. 6: High-level computing using FE-PS-based array.

a Schematic circuit diagrams of a 1 × 2 FE-PS-based array for demonstrating in-sensor MAC operation. b, c Measured output photovoltages of the 1 × 2 array (i.e., V_ph-1&2) at t_ill ~ 20 s when different combinations of I_light-1 and I_light-2 are applied to the two FE-PSs in the array. In (b), the two FE-PSs are set in the same state with R_pv-1 = R_pv-2 = 0.048 V·cm²·mW^-1 and V_offset-1 = V_offset-2 = 0.7 V, while in (c), they are set in the opposite states with −R_pv-1 = R_pv-2 = 0.048 V·cm²·mW^-1 and −V_offset-1 = V_offset-2 = 0.7 V. The theoretical values of V_ph-1&2 are indicated by the star symbols. d Schematic diagram of a 4 × 9 FE-PS-based array for the recognition of letters “S”, “C”, “N”, and “U”. e Expected and f measured V_ph values (@ t_ill ~ 20 s) of the FE-PSs after programming, which are arranged into 4 groups with each connected to a specific output neuron. The expected V_ph values are obtained from the trained weights. The weights are first trained ex situ, then binarized to 1 and −1, and eventually programmed into the array. After programming, the weight 1 (−1) corresponds to an expected V_ph value (@ t_ill ~ 20 s) of 2.4 (−2.4) V under the I_light of 70 mW/cm². g Temporal output V_phs at different neurons when the letter “C” is presented to the array. h Summary of the measured output V_phs (@ t_ill ~ 20 s) at different neurons for different input letters. The corresponding expected V_ph values are shown in the brackets.