Document Type
Poster
Date of Original Version
3-27-2026
Abstract
This work discusses how to directly extract features from complex-valued spectra using an energy-based probability distribution that represents complex-valued data. As an extension of the well-known restricted Boltzmann machine (RBM), the proposed model, the complex-valued restricted Boltzmann machine (CRBM), is made to deal with complex valued visible units. With the use of contrastive divergence (CD) or Gibbs sampling, the CRBM, like the RBM, develops a grasp of the relationships between visible and hidden units without requiring connections between units in the same layer. We propose a new in-sensor computing architecture that can dramatically reduce the amount of data transferred from the sensing board to the DRAM of the main computing system, giving rise to improved performance and energy efficiency. The sensing board investigated here is mmWave (millimeter Wave) radar that generates and transfers the Inphase(I) and Quadrature(Q) data to the central ADAS (Advanced Driver-Assistance System) computing platform for object detection and recognition in autonomous driving systems. Rather than transferring raw I/Q data, our new architecture performs initial stages of machine learning computation on the sensor board and transfer partial results to ADAS computing system, leading to decreased latency and energy consumption while still maintaining the data accuracy. We present an energy based probabilistic graphical model, CRBM (Complex-valued Restricted Boltzmann Machine), a generative model to reduce the raw data size to be transferred to the ADAS computing platform.
Recommended Citation
Karakus, Erkan and Yang, Qing, "Machine Learning in Sensors: Feature Learning from Complex Spectra in mmWave Radar Signals" (2026). Poster Presentations. Paper 23.
https://digitalcommons.uri.edu/gradcon2026-posters/23