Fine grained analysis
Fine-grained categorization refers to the problem of distinguishing between closely related entities, e.g., Monarch butterfly vs. Viceroy butterfly based on visual features, or Louisiana Waterthrush vs. Swainson’s Warbler based on auditory features. Analogous problems exist within medical data, language sentiment analysis, and remote sensing. Making progress in fine-grained categorization requires a combination of advanced ML research and coordination with – and respect for – human communities of knowledge.
Based on mathematical and statistical modeling of visual representations and on large scale experimental work this collaboratory will make foundational contributions to the centre’s basic research areas:
Explainability: Explore the fundamental question: How can AI assist human decision making in fine-grained medical domains? Construct AI systems for fine-grained classification that are causally explainable and interpretable by humans.
Self-supervised learning: Push the limits to fine-grained classification by self-supervised learning. Based on expert knowledge we will exploit masked learning to establish representations for fine-grained few shot generalization.
Novelty detection: Develop novelty detection at scale to cope with extreme class imbalance and other effects of long-tailed distributions in real-world fine-grained classification problems.