Our next generation of transport does not have the freedom to consume the earth's resources that previous generations have.In this seminar we consider the impact that environmental constraints are having on the vehicles in design right now and how some of the challenges are being met by individual vehicles communicating to each other and back end systems.

Since the 1980s, the object of design for dependability has been to avoid, detect or tolerate system faults so that these do not result in failures that are detectable outside the system. Whilst this is potentially achievable in medium size systems that are controlled by a single organisations, it is now practically impossible to achieve in large-scale systems of systems where different parts of the system are owned and controlled by different organisations. Therefore, we must accept the inevitability of failure and re-orient our system design strategies to recover from those failures at minimal cost and as quickly as possible. This talk will discuss why such recovery strategies cannot be purely technical but must be socio-technical in nature and argue that design for recovery will require a better understanding of how people recover from failure and the information they need during that recovery process. I will argue that supporting recovery should be a fundamental design objective of systems and explore what this means for current approaches to large-scale systems design.

The most successful attempts to merge logic and neural networks have formed two disciplines in AI. One field - called statistical relational learning - is focused on the statistical data, while considering the logical (or relational) knowledge as a meta property arising from the statistical or probabilistic models. The second field - neuro-symbolic integration - comes to the problem from the point of view of formal logic, and looks for machine learning methods or neurocomputing methods that can be useful for the purposes of logical inference.

It is common for neuro-symbolic systems to propose unconventional models of neural networks to solve conventional logic problems. In contrast, the statistical relational learning would generally tend to use conventional learning methods, while the relational (or logic) level of reasoning would be somewhat different from the mainstream methods used in computational logic.

In this talk, I will consider several examples of recently proposed neuro-symbolic systems, and analyse their capacity to do either logic inference or learning (or both) along the lines above.