Real-time Analog Processing with On-Chip Learning using Multiple-Input Translinear Elements

Abstract Analog processing has re-emerged as a mode of computing complex real-time algorithms since it consumes less power. Analog circuits are susceptible to mismatches, and variations could be mitigated with onboard training. In this study, we are presenting the use of analog floating-gate MITE-based circuits for various computations. The study showcases their use in performing various basic arithmetic operations. Analog computation can provide a significant advantage in applications where signals are in the analog domain. Here, the study uses a neural decoding task to demonstrate an application where input neural data are mapped to kinematics. The study demonstrates a real-time neural decoding task with an analog adaptive circuit. Additionally, we develop an on-chip learning algorithm to adapt the parameters of the analog adaptive circuit. For a neural decoding task, on-chip learning enables the improvement of the overall Pearson correlation coefficient from 0.07 to 0.69. On-chip learning and adaptation can significantly reduce the need for off-chip communication in implantable devices.