FPGA-based Asynchronous Spike Perceptron for Tiny MEMS Tactile Sensors

We propose an asynchronous spike perceptron (A-SP) for a micro-electromechanical system (MEMS) tactile sensor on a field-programmable gate array (FPGA) and constructed a multi-layer perceptron (MLP) model configured with our A-SP (Multi-Layer A-SP) and evaluated it for an XOR neural network task. Our A-SP uses digital circuits to process the data from tactile sensors, so it is faster and consumes less power than a traditional analog circuit. The results of Multi-Layer A-SP indicate that our A-SP worked well for a typical 2-input, 1-output XOR. The power consumption increased by 8.6 mW for each additional MLP model. We also constructed an echo state network (ESN) configured with our A-SP and evaluated it. The results indicate that the A-SP-based ESN accurately generated XOR outputs from one time-series signal. These results indicate the effectiveness of constructing a more complex recurrent artificial-intelligence model for processing time-series sensor data.