Skip to main content


The development of Deep Learning techniques and algorithms for radio systems is currently under the spotlight, carrying the promise of solutions to problems once considered too difficult, too complex or too inflexible. While for some problems Deep Learning has not shown not to be able exceed the performance of previous traditional algorithms based on analytical findings, for others, such as radio fingerprinting, spectrum usage detection and classification, indoor location, transmitter-receiver constellation adaptation, Deep Learning pro vides a cost-effective way to deal with these problems. However, one of Deep Learning's biggest advantages is also its largest Achilles heels: the data used to train it. Deep Learning, relies on a large amounts of data to train its also large number of parameters, which is essential to making the Deep Learning system actually learn what is intended of it as well as generalizable to different environments. A supervised learning system can be easily fooled by biased datasets, providing good classification results while basing its decision on bad criteria.

This project proposal aims to build on previous experience on training supervised and unsupervised Deep Learning systems for radio to enhance Deep Learning's performance through better algorithmic construction and better dataset creation. We will exploit the power of the CorteXlab experimental radio testbed, powered by its 42 Software Defined Radios (single and multiple-antenna), mobility robot and isolated radio environment. One characteristic unique to CorteXlab is its reproducible experimental environment and capabilities, which are key to enabling fair comparison among techniques to assess correct performance evaluation. 

Initially, the problems of radio fingerprinting, indoor localization and spectrum usage detection and classification problems will be addressed, but the project can bed extended to a number of physical layer techniques that pose similar requisites.