Keynote Speaker

Prof. Jason Ralph University of Liverpool

Instabilities in Navigation - Balancing on the head of a pin

The problem of navigation with dead reckoning is one of instability. Many of the usual methods that we learn from school and college are problematic as a result. We can augment dead reckoning using alternatives that provide direct measurements of position, but none are perfect. This talk outlines the main problems associated with dead reckoning navigation and the use of inertial sensors to measure the dynamics of a vehicle/platform. We discuss current sensor technologies and the possible use of quantum technology to augment classical inertial sensing, and we try to highlight the difficulties in using signal processing in unstable systems of equations. 

Machine learning (ML) has been successfully applied to a very wide range of defence signal processing and communications problems. A few key challenges deserve further attention. First, there often is a lack of sufficient signals/data to train the ML algorithms in use. Second, the huge of volume of signals/data are often collected by sensors at geographically distributed locations. Third, after proper training, the trained models may operate in an environment different from that where the training signals/data is collected. This talk will present exemplary techniques to address these challenges and briefly discuss open issues for future work. 

First, we consider a classification problem of electromagnetic (EM) signals to illustrate a technique to overcome lack of training data. Specifically, a system using Generative Adversarial Network (GAN) will be presented that can detect and classify EM signals as friendly or hostile, even when there is no prior data of the hostile signals. The proposed approach is validated by use of a simulated waveform dataset. Second, to support defence applications, federated learning can be used to learn the model parameters from signals/data collected at distributed nodes, without data sharing with any other node, and adapts according to the limited availability of resources. Using real datasets, the experimentation results show that the proposed approach performs near to the optimum with various ML models. Third, by using a network for defence analytic processing, we highlight the potential advantage of transfer learning for speeding up reinforcement learning when the operating environment has significant changes. 

Dr Alexandrou Serb, University of Edinburgh

Adiabatic computing for low power image sensing

The world has an insatiable appetite for data, which leads to vast demand for data processing. Nowhere do the constraints of this model -especially power constraints- become more evident, than in the domain of image processing. We at the univ. of Edinburgh are engineering a hardware accelerator aimed at making neural network-based image processing more power efficient. Crucially, the accelerator uses “adiabatic techniques” -which will be explained in the talk- in order to reduce the power consumed by the neural network it embodies to below the CV² limit typically associated with digital circuits, and which forms a hard limit for non-adiabatic systems. This is significant, as 90%+ power dissipation savings can be made according to data gathered so far, allowing designs implemented in e.g. 180nm CMOS technology to be theoretically competitive with designs in 65nm CMOS (which also has strategic implications given the current state of semiconductor markets). In this talk we will cover the basic experiments and results carried out so far in this nascent field and discuss the perspectives for the future development of the technology.