AI@CIS_StrathclydeOur research

We carry out a range of research activities in human centric AI, including:

Human Centric AI for Healthcare

We envision a new research direction in AI for health to foster a transformative human centric paradigm, in which an emerging collaborative relationship between computers and humans will be created based on the ever-increasing power of AI. The computers will no longer just be used as tools, but instead act actively as a co-worker to offer sound, clear and evidence-based solutions to support clinicians and patients in clinical decision making and enhance clinical outcome.

The research focus on three specific research themes:

  • Trusted Data: Uncertainty, risk and security aware modelling of health data that will handle randomly varying nonlinearities, uncertain and missing measurements, where new approaches to estimate risks and predict outcomes of clinical decision support for informed choice in the face of risk and uncertainty will be targeted.
  • Trusted AI: Explainable and secure AI that will contribute to the trust in the clinical decisions recommended by AI by investigating innovative AI-based health solutions that are explainable, secure, trustworthy and acceptable by humans.
  • Trust in Human Factors: Human centred AI design that will investigate human factors and identify barriers in clinical acceptance to support a transformative human & AI collaboration experience based on trusted AI and data in the context of healthcare.

AI & Software Engineering

We are looking at how artificial intelligence techniques can be used to support the process of software engineering with the aim of creating systems that are robust, reliable and adaptable. We are also looking at the converse problem of how AI systems should be engineered, which is particularly important as AI technologies are becoming commonly deployed in critical situations. Research themes within this topic include:

  • Test Data Generation: For some time now, it has been established that AI search-based strategies are capable of generating effective sets of test data for a system, freeing the software engineer from this tedious task. But there are still real challenges such as generating test data that is meaningful to the engineer, and testing non-traditional systems (such as AI systems themselves).
  • Test Outcome Classification: While it is possible to generate large volumes of test data, the output from the system under test still needs to be checked. To this end we are investigating how AI-based techniques such as anomaly detection can be used to effectively distinguish between passing and failing tests.
  • Fault Localisation: If we detect a failing test, how do we identify the associated code that needs to be fixed? Approaches under investigation here involve either looking at the different paths through the system taken by passing and failing tests, or using text-based analysis techniques to work from user-supplied bug reports.
  • Autonomous Systems Evolution: Once we have detected a fault, how do we go about fixing it? Or how can we rapidly respond to changes in requirements, patterns of systems usage, or environmental changes? Search-based AI techniques have the potential to evolve code and effectively search through a huge number of potential solutions to generate fixes or new system configurations that autonomously adapt to internal errors, external environmental changes, and previously unseen (or unimagined!) scenarios.

AI & Animal Health

 This includes research in:

  • Interpreting novel sensor/IoT sources: improved surveillance and feedback for disease control
  • Machine Learning in noisy environments: biology is ‘messy’ and model validation hard
  • Agent-based modelling: treatment impacts and cost-benefit trade-offs across control strategies
  • M-Health in LIMC: embedding AI-based models to increase diagnostic competence in the field

 

AI & Video Coding, National Grid and E-Learning

This includes research in:

  • National Grid ESO: An AI solution to generate the optimal switching sequence for network topology running arrangements
  • FAIR E-LEARNING: An ML and video coding solution to students’ accessibility to online video learning content to both privileged and under-privileged groups
  • Video Coding for Machines: Video perception by machines is different from humans
  • New concepts in video coding for machines will be developed to optimise the bandwidth and other resources
  • AI and ML for real-time video streaming in automated driving applications: high quality real-time video streaming is a key in automated navigating in emerging smart transport applications
  • An intelligent video streaming platform will be developed for automatic navigating of trucks in ports

AI & Multimodal Speech Communication

Speech and emotion are multimodal signals, meaning that both production and understanding make use of a variety of cues, including audio speech information, mouth movements, other facial expression changes, environmental information such as background noise and other contextual cues. We are investigating feature-based approaches in partnership with machine learning to produce communications systems. This includes:

  • Lip reading and speech recognition
  • investigating the use of machine learning and fusion of the audio and visual streams
  • Explainable image features, with a particular focus on the use of Gabor wavelets to extract meaningful temporal mouth parameters
  • Emotion recognition from video data, using both machine learning models and feature based methods to analyse and classify facial changes

 

Robust Reasoning for Intelligent Agents

We are working on the intersection of machine learning and task planning with the goal of developing long-life autonomous systems that are robust and trusted. As these systems become more complex, they become more opaque to non-expert users. As a result, current applications of autonomous systems in challenging environments, or interacting with untrained, non-expert users are limited to operating within environments that have been adapted to support them, very small time windows of operation, or the behaviour of the system is tightly confined. Our vision is to develop novel approaches to intelligent control that are capable of reacting robustly and safely in dynamic and challenging environments, explaining their behaviour, and working within mixed teams of humans and machines.