SAR object classification implementation for embedded platforms

TL;DR: An algorithm to reduce the run-time and the power consumption of Deep Neural Networks (DNNs) classifiers by reducing the DNN model size required for a given object classification task is developed.

Abstract: This research details a new approach to optimize neural network architectures for Synthetic Aperture Radar (SAR) object classification on neuromorphic (e.g., IBM’s TrueNorth) and embedded platforms. We developed an algorithm to reduce the run-time and the power consumption of Deep Neural Networks (DNNs) classifiers by reducing the DNN model size required for a given object classification task. Reducing the model size reduces the number of mathematical operations performed, and the memory required, enabling computation on low size, weight and power (SWaP) hardware. We will provide our approach and results on relevant SAR data. Our entirely new approach starts with a very small multi-class convolution neural network (CNN) and replaces the standard negative log likelihood loss function with a single-class log loss function. We then generate an ensemble of small models trained for an individual class by varying the training data using a k-fold cross-validation and augmentation. This is done for each class and the resulting ensembles classify objects by finding the maximum average probability across each ensemble of single-class classifiers. We demonstrate 91-99 percent classification accuracy on three different datasets with composite networks that require almost 10 times fewer mathematical operations than SqueezeNet (a reduced parameter CNN with AlexNet performance).