Bio-Inspired Neural Group Testing for Scalable Inference in Sparse Radar Detection

This paper presents a novel neural group testing method to detect sparse radar signals over 3.5GHz Citizens Broadband Radio Service (CBRS) band with low latency and computation complexity. Motivated by recent advances in neural group testing applied to images, we explore various approaches inspired by COVID-19 testing strategies such as pooling and sparse recovery, including multi-round tree merge and our own novel hybrid method. Our proposed method demonstrates up to a 51% increase in speed compared to individual testing at a 0.25% prevalence of radar signals in the dataset. It integrates the strengths of both least absolute shrinkage and selection operator (LASSO), which requires fewer tests and offers parallelizability with the simplicity of implementation needing only one network, and the binary tree method, which efficiently reduces the number of samples and eliminates many negative samples in its initial step. A custom sounding reference signal (SRS) dataset, captured using USRP X310, is combined with the sparse radar data from the NIST 3.5 GHz CBRS Band dataset to validate our approach.