Postgraduate research opportunities Neuromorphic RF sensing and processing

Artificial Neural Networks (ANN) have been a research topic since the 1940s, even though it was only after the formalisation of the backpropagation algorithm for ANNs in 1986, that it became possible to effectively train a non-linear ANN with multiple hidden layers of neurons. Nowadays, deep Neural Networks (DNNs), such as CNNs, as opposed to shallow ANNs, are state-of-the-art, not only in image processing for object detection but in almost every pattern recognition application, so that machine learning using DNNs has been de facto rebranded as Deep Learning (DL). The DL community has understood since its inception in the early 2000s that hardware acceleration of DL architectures was an essential part of modern CNNs. However, there is a general consensus that in the future Von Neumann-based computer architectures will not be able to cope with the computational demand imposed by modern DL structures.

In this context, the concept of neuromorphic (NM) technologies provides an attractive alternative to conventional computing architectures. The key characteristic of neuromorphic systems, as opposed to conventional ANNs and CNNs, is not in the type of physical implementation, but rather in the fact that NM systems generate and propagate spikes as a means of processing data. The information is indeed encoded in the timing and rate of spikes, generated by each neuron in the network once a set of incoming stimuli crosses an inner threshold within the neuron. To highlight this difference between ANNs and CNNs, it is said that NM systems use spiking neural networks (SNN). Therefore, truly neuromorphic systems are those that, either through an analog or digital implementation, make use of spikes, rather than numerical values, to propagate information through the network.

Some of the key advantages of SNNs over DNNs include higher efficiency and speed and a significantly lower power consumption requirement.

While neuromorphic sensing and processing have attracted interest from the electro-optics community, the potential that this technology has for the lower range of the electromagnetic spectrum has not been investigated yet. In particular, RF sensors (active and passive) exploiting the neuromorphic paradigm have not been investigated as well as the use of SNNs to process the data obtained by this family of sensors.

Current RF sensing systems, either active or passive, rely on classic architectures in hardware and processing; for example, the RF components and architectures of existing radars are not much different in principle from the first radars developed during WWII. At the same time, the processing philosophy also relies on classic signal processing such as detection and estimation theory. Neuromorphic offers the opportunity to innovate in both hardware and software approaches in radar sensors, from developing novel brain-inspired approaches to sense the RF environment, to the use of SNN to exploit the sparse characteristics of received radar signals, as well as allowing novel encoding paradigms for the relevant properties of the detected RF signals. In addition, for RF sensors (ie Electronic Surveillance Systems)  the exploitation of neuromorphic paradigms can help to address open challenges such as the receiver bandwidth vs data representation vs heat generation on board problem of modern Ultra Wide Band RF receivers.

The aim of this project will be to investigate neuromorphic solutions in RF sensing at both hardware and software levels. The main objectives of the project will be:

• to become familiar with fundamental concepts of radar principles and RF sensing systems.
• to become familiar with fundamental concepts in the field of AI and ML, with a particular interest in sensing systems.
• to identify constraints and training requirements for SNNs when dealing with RF data;
• to identify and design brain-inspired passive RF sensing modes to generate spiking outputs;
• to design and develop novel NM/SNN frameworks, for RF systems, based on simulated and real data.
• to validate the developed framework on real data provided by Leonardo and acquired at the SSP&S labs.
• to devise a testing strategy to evaluate and demonstrate the performance of the novel NM/SNN models and techniques developed, and compare performance with traditional processing methods.
• to investigate the experimentation with Neuromorphic hardware of the developed solutions;
• to disseminate outcomes from this research, in world leading journals and conferences.