Departmental Seminar: “Trustworthy AI for Subtle Visual Signals” – Dr Xinqi Fan, Manchester Metropolitan University

About

Abstract Some of the informative visual evidence appears as small changes that are easy to overlook. Learning from these subtle signals remains challenging for AI systems. This talk presents our research on trustworthy AI for recognising subtle visual signals across facial and medical image understanding. First, facial micro-expressions are brief and subtle movements that may provide cues to underlying affective states. We investigate robust representation learning through self-supervised motion learning to capture their fine-grained dynamics. We further demonstrate that the learned representations can support both micro-expression analysis and video generation. Second, we turn to medical imaging for ulcerative colitis, where small differences in mucosal appearance can affect disease assessment. To ground predictions in clinical knowledge, we develop an endoscopic-informed spiral-scanning strategy for state-space models that captures meaningful visual patterns. We also introduce a mixture of low-rank vision-language experts to incorporate clinical concepts into model training and reasoning. Finally, small differences between training and deployment conditions can reduce model performance. We address this distribution shift through retrieval-augmented test-time adaptation and reflective multi-agent reasoning, enabling AI systems to use new evidence and self-improve their predictions without conventional retraining.