All-Analog Silicon Integration of Image Sensor and Neural Computing Engine for Image Classification

We have designed a fully-integrated analog CMOS cognitive image sensor based on two-layer artificial neural network and targeted to low-resolution classification. used single poly 180 nm process technology, which includes modules for realizing the building blocks of sensor. Our design all sub-circuits required perform sensing task, from output classification decision. The weights are stored in single-poly floating-gate memory cells, using transistor per weight. This enables classifier be intrinsically reconfigurable, trained various problems, images. As case study, capability is tested version MNIST dataset handwritten digits. circuit exhibits accuracy 87.8%, that comparable an equivalent software implementation operating digital domain with floating point data precision, average energy consumption 6 nJ inference, latency 22.5 $\mu \text{s}$ throughput up 133.3 thousand inferences second.