Adaptive Pre-Processing for Neural Network Hardware Deployment

Neural Networks are gaining popularity in the signal processing (SP) field. In radar SP, micro-Doppler based classification with Convolutional Neural Networks has shown state-of-the-art results. Still few works focus on top-down hardware deployment, and ID spectrum-based classification approaches with wide amplitude dynamic lose significant accuracy during quantization. In this paper, the problem is tackled with an adaptive pre-processing strategy that is able to mitigate accuracy loss upfront, learning the optimal representation of the data based on the classifier loss. To show this, a Drone vs. Unknown spectral classification scenario is presented.