Strands of AI research
Researchers in the School of Informatics explore a number of different strands of artificial intelligence research.
One strand of AI research concerns perception, where we wish to understand signals in the world such as images, video and speech. For speech, the aim is to turn the acoustic signal into words (research carried out by Centre for Speech Technology), while for vision we may wish to identify objects, agents and their actions (Bob Fisher, VICO, Laura Sevilla, Machine Intelligence Group). We can also combine modalities, e.g. with the joint analysis of speech and video data.
As well as perceiving the world, we may want to reason and answer questions about it. Logic provides a powerful framework for this task, as applied e.g. to mathematical theorem proving (Mathematical Reasoning Group).
Another important aspect is being able to reason properly under uncertainty when there is incomplete information, for example in a medical application where there may be multiple possible diagnoses, some more likely than others (Machine Learning Group).
Artificial intelligence systems need to adapt and learn (Machine Learning Group). In supervised machine learning, labelled data is provided, along with an evaluation of how well the task is being achieved, and the system adapts to improve its performance. For example, a system can be trained to label the objects in an image, like a bird or a car. Such problems have benefited from recent developments in deep learning. We also develop algorithms for unsupervised learning, to detect such structure in the data without the benefit of labels.
We build agents that can act in their environment to achieve useful tasks, be it the physical world for robots, or virtual worlds for softbots. Such agents need to learn, reason and plan, whether they are acting fully autonomously, or in collaboration with humans and computer-based entities. Controlling dextrous motion and physical interaction (Michael Mistry, Zhibin Li) provides the basis for intelligent action and control of robots (Statistical Machine Learning and Motor Control Group, Steve Tonneau). Generating such actions can also be applied to virtual agents in computer graphics and animation (Computer Graphics and Visualization Research Group). Concepts of autonomous decision making can focus on the coordination and cooperation in multi-agent systems (Autonomous Agents Research Group), and be extended to other domains such as predictive modelling and decision making in energy and environmental systems (Robust Autonomy and Decisions Group).
Language is a critical route by which we communicate and convey information to each other. We work on algorithms that understand the structure of sentences (The Cohort research group), determine the meaning of the linguistic input, and then generating appropriate responses (Natural Language Processing Group). Natural language understanding and generation can be deployed in range of areas; particular strengths in Edinburgh include machine translation (Statistical Machine Translation Group), document summarization (Extreme Summarization), and information retrieval, including from social media (Walid Magdy).
The study of natural systems can provide useful information for the development of AI, and vice versa. This can be at the level of computational neuroscience (Computational Neuroscience and Neuroinformatics) which studies how the brain processes information; inspiration for behavioural control from insects (Insect Robotics Group); computational psychiatry, which uses reinforcement learning models to understand the factors behind individual variability and clinical disorders (Peggy Series); and behavioural experiments with human subjects which can be used to underpin computational cognitive models (Joint Eyetracking Lab).
Safety, trust, ethics of AI
It is increasingly important that designers of AI systems can provide assurances regarding properties like safety, trustworthiness, ethics and fairness. Rigorous computer-based reasoning about the agents’ models and algorithms is one route to help address these concerns (Vaishak Belle, Mathematical Reasoning Group). In order to build AI systems that people actually find useful, we need to go beyond purely technological solutions. Design Informatics (Design Informatics) connects AI specialists with designers, cognitive scientists, and human-computer interaction experts to create fair, safe, user-centric AI applications.
Researchers in the School develop applications of AI-based systems in a wide variety of areas. In addition to those mentioned above, these include: modelling care pathways in healthcare (Workflow FM); quantitative modelling and real-time data analysis techniques for finance and business (Tiejun Ma); and using transfer learning to study cancer evolution (Guido Sanguinetti).