A Gradient-Descent Calibration Method to Mitigate Process Variations in Analog Synapse Arrays

A technique to compensate the effects of process variation in analog-based neural computation circuit arrays is presented. The proposed technique considers the process variation that degrades the accuracy of a neural network as a variable for optimization and adjusts the circuit parameters via a gradient descent method. In the TensorFlow framework, the effects of process variation on the neural network are modeled and the improvement in the inference accuracy achieved by the proposed technique is analyzed. For a 4-layer CNN MNIST example driven for 1000 different variations sets, the technique restores the accuracy to the original level, which can degrade to 17.3~95.2% due to process variations.