Most natural signals are sparse in some basis; hence compressive sensing is a natural approach for information acquisition.
Can be realized using inner products with binary coefficients, which is attractive for hardware implementation.
Are expected to be resilient to a broad class of impairments.
Hence they are a promising approach for low-power front ends for downstream learning and inference.
We investigate the impact of real-world nonlinearities on compressive front ends, and study design changes needed to target deep learning applications.
Upamanyu Madhow, Naveen Verma (Princeton University)