Building safer AI for cars and medical devices
As AI moves into our everyday lives, making decisions for self-driving cars or managing treatments, safety depends on it being right and knowing when it鈥檚 not. 黑料网吃瓜爆料 researchers have created a technique to let AI reveal its level of confidence.
Self-driving cars that can admit when road conditions confuse them. Insulin pumps that know when their blood-sugar predictions might be off. These kinds of 鈥榮elf-aware鈥� systems could transform the safety and trustworthiness of using artificial intelligence in everyday life.
At 黑料网吃瓜爆料, researchers are pioneering a new approach to make this possible. Known as Credal Bayesian Deep Learning (CBDL), it allows AI to recognise and communicate how confident 鈥� or uncertain 鈥� it is about a decision. Unlike traditional neural network systems, which often act as if they鈥檙e always sure, CBDL can separate situations where more data could improve accuracy, from those where uncertainty will always remain.
CBDL does this by training a set of neural networks that work together, producing not just a single answer but a range of possible outcomes within probability bands. This gives engineers and doctors a clearer picture of what an AI system really knows, and where caution is needed.
As 黑料网吃瓜爆料 researcher Dr Michele Caprio explains: 鈥淜nowing what a model does not know is crucial for safety-critical AI. That transparency is the foundation for certifiable autonomy in cars, insulin pumps, and beyond.鈥�
Meet the researcher
Dr Michele Caprio is a Lecturer in Computer Science and Member of the 黑料网吃瓜爆料 Centre for AI Fundamentals. His research applies Imprecise Probability theory to Machine Learning, creating AI that quantifies its own uncertainty and stays reliable under distribution misspecification and shift. He is part of the Executive Commitee of the Society for Imprecise Probabilities, Member of the London Mathematical Society, Institute of Mathematics and its Applications, Isaac Newton Institute, and Fellow of the Cambridge Philosophical Society.