CAREER: Learning to Sense

Motivation

Conventional sensing systems acquire data in a fixed way, independent of the task the data will actually serve, and the acquisition, reconstruction, and inference stages of the processing pipeline are designed and optimized separately. The result is an ever-growing volume of measurements and an unsustainable demand for power, storage, processing, and communication.

Approach

This project develops an adaptive, task-oriented and physics-aware data-to-decision pipeline that jointly optimizes what is measured, how it is reconstructed, and how decisions are made — all within a single learning framework. Work is organized along three connected directions:

• Physics-aware learned reconstruction — network architectures derived from regularized inverse problems and classical signal processing models, so that what the network does remains interpretable.
• Task-oriented measurement design — learning constrained, hardware-realizable measurement operators jointly with the reconstruction and inference blocks that follow them.
• Direct inference from few measurements — classification and parameter estimation performed straight from a small number of learned measurements, without a full reconstruction step.

Selected outcomes

• PLFNets: Interpretable Complex-Valued Parameterized Learnable Filters for Computationally Efficient RF Classification, IEEE Transactions on Radar Systems, 2024.
• CV-SincNet: Learning Complex Sinc Filters From Raw Radar Data for Computationally Efficient Human Motion Recognition, IEEE Transactions on Radar Systems, 2023.
• HRSpecNET: Deep Learning-Based High-Resolution Radar Micro-Doppler Signature Reconstruction, IEEE Transactions on Radar Systems, 2024.
• Learning-Based Optimization of Hyperspectral Band Selection for Classification, Remote Sensing, 2023.

The project also supports outreach and educational activities that introduce K–12 and university students to sensing systems, signal processing, and machine learning.