Synopsis: Modern autonomous mechanical systems are saturated with sensors. Faults that occur due to false sensing or runtime errors as well as hardware failures need to be detected quickly and the root cause component that caused that failure must be diagnosed and fixed. This presents significant challenges: 1) quickly detect the fault with high precision, 2) identify the root cause of the failure (diagnosis), 3) support a decision which is derived from the condition of the component. We present a novel AI paradigm for fault detection, diagnosis, prognosis and troubleshooting of faults in autonomous mechanical systems. We demonstrate it on cars, but the AI approach is general.

Video Book meeting here

Synopsis: AI can have enormous impact on many technologies. In our talk we will focus on two areas that in recent years have experienced the increased use of AI: Autonomous Vehicles (AVs) and city-scale Digital Twins (DTs). Gothenburg is an exciting hub for the development of these technologies and Chalmers Industriteknik an important interface to facilitate uptake of AVs and DTs. We will present some recent initiatives and research centres located in Gothenburg and discuss what influence AI can have on the development of these technologies. DTs must be based on consistent and accurate 3D City Models. AI can help us to generate and enrich DTs of cities in many ways. For AVs the use of AI and Machine Learning centers around challenges such as perception, semantic segmentation and lane keeping.

Video Book meeting here

Synopsis: Accurately analyzing and predicting movements of objects through time and space is central to many applications. We do research on learning generative models based on trajectory data, probabilistic logical reasoning over observed and predicted trajectories (and other time series data), and privacy-preserving synthetic data generation. We mainly work with Gaussian Processes and temporal GANs (Generative Adversarial Networks) for learning generative models, probabilistic signal temporal logic for reasoning, and Bayesian Optimization for synthetic data generation. Our research is applied to both autonomous systems, such as unmanned ground and aerial vehicles, and traffic monitoring and analysis together with for example the Swedish Transport Administration (Trafikverket) and the local transportation authority.

Video Book meeting here

Synopsis: How to do machine learning if you do not have good enough data available for your application? The answer is semi-supervised learning – given that you do not have unlimited (financial) resources, as annotation ('labeling') of data is often too much of an investment to make. Can we do something else? By working with partially labeled data, having some data samples getting full annotation, but most data remaining unlabeled ('unsupervised'), better data can be secured. This strategy is as promising to industry as it ever was. Semi-supervised learning touches on topics like anomaly detection, active learning and others, and should be of interest to many emerging players innovating in the AI space. The problem of designing good algorithms for semi-supervised learning is challenging though, and inexpert attempts may result in biased or even unreliable results. This talk will give an introduction to the state-of-affairs of semi-supervised learning.

Video Book meeting here

Synopsis: Video is recorded everywhere, but its usefulness can be enhanced substantially by (i) careful use of audio. On one aspect we can enhance speech of a visible speaker by removing sounds unrelated to mouth movements. On another hand we propose to remove speech altogether in order to perform audio scene analysis while complying with privacy laws. (ii) Detection of anomalies, which are the most important events recorded in videos and by other sensors.

Video Book meeting here