Fig. 1. Working principle of bipolar cells in the mammalian retina

Fig. 2. Schematic diagrams of bipolar cell-inspired MSM structure unit device and its array configuration

Fig. 3. Energy band distributions of the BP-InAs-BP structure in the Sentaurus TCAD physical simulation. (a) Dark state; (b) light state

Fig. 4. Schematic diagrams of different structures and corresponding leakage currents in dark state. (a) Structure of BP-InAs-BP; (b) leakage currents in dark state of different structures; (c) structure of heterojunction; (d) structure of pure BP; (e) structure of pure InAs

Fig. 5. Asymmetric distributions of cross-section in traditional BP-InAs heterojunction MSM phototransistors. (a) Space charge distributions under positive and negative biases at 832 nm wavelength; (b) corresponding band distribution

Fig. 6. Symmetrical distributions of the cross-section in BP-InAs-BP MSM phototransistors. (a) Space charge distributions under positive and negative biases at 832 nm wavelength; (b) corresponding band distribution

Fig. 7. Schematic diagrams of different structures and corresponding electron current density distributions under negative bias in the optical state. (a) Structure of BP-InAs-BP; (b) structure of heterojunction; (c) structure of pure BP; (d) structure of pure InAs

Fig. 8. Device prepared in the laboratory and its simulated fitting results. (a) Optical microscope image of the BP-InAs-BP MSM photodetector; (b) comparison of experimental data and model test data of the proposed device in the dark state and 832 nm infrared light

Fig. 9. Responsivities under various biases at 832 nm wavelength

Fig. 10. Current responses under different light intensities at 832 nm wavelength

Fig. 11. Distributions of absorption density and light-excited carriers at the InAs-BP heterojunction in proposed device at different light wavelengths. (a) 520 nm; (b) 832 nm; (c) 1310 nm

Fig. 12. Schematic diagram of the convolutional neural network implemented by proposed array for modified national institute of standards and technology database (MNIST) recognition

Fig. 13. Comparison of recognition rates between the fully connected in the proposed arrayed convolutional layer and traditional fully connected

Fig. 14. Confusion matrix of the recognition accuracy after introducnig the proposed arrayed convolutional layer