Nonlinearity Modeling for Mixed-Signal Inference Accelerators in Training Frameworks

In this work, an approach for modeling the nonlinearity of a mixed-signal neural-network accelerator in training frameworks is presented. We extend the state-of-the-art by modeling a mixed-signal neuron in a neural-network training framework such as TensorFlow. It is shown how the nonlinearity can be integrated in the anyhow required quantizer models. The parameters of the nonlinearity model of a single neuron are found by a preliminary training, where the model variables are treated as learnable parameters, while the behavior of the modeled neuron is fitted to circuit-simulation or -test data. The model is never moved to another toolchain and the entire model extraction process and the process of training a neural network under the influence of circuit-nonlinearities happen in the training framework, where TensorFlow is chosen for this work. We evaluate the approach by analyzing how a full-scale VGG-16 based CIFAR-10 classifier adapts a known neuron nonlinearity. The impact of the nonlinearities can be removed by training and without performing improvements on circuit level.