Design of Analog Computers: Building Blocks and Advanced Methods for Tiny Analog Machine Learning with Surrogate Modeling

This study investigates the use of surrogate modeling of electronic circuits for functional optimization problems using gradient-based methods, demonstrated with an Analog Artificial Neural Network use-case. The results show that a surrogate model approach is more computationally expensive than a pure mathematical model (about 10 times), that the training process using error gradient back-propagation is instable and only randomly convergent, but it significantly improved classification error shown with the IRIS benchmark dataset, with the adam optimizer finding the "golden standard" solution with 2% classification error. The training results can be directly transferred to an analog transistor circuit without loss of accuracy.