Robust Mixture Learning When Outliers Overwhelm Small Groups

Watch a 42-minute research lecture from the Joint IFML/MPG Symposium where ETH Zurich's Fanny Yang explores the challenges and solutions in estimating means of well-separated mixtures under adversarial conditions. Delve into the novel concept of list-decodable mixture learning (LD-ML), particularly focusing on scenarios where outliers outnumber smaller cluster groups. Learn about a groundbreaking algorithm that achieves optimal error guarantees while minimizing list-size overhead, surpassing existing list-decodable mean estimation methods. Understand how this approach excels particularly in separated mixture scenarios by leveraging mixture structure for partial sample clustering before applying iterative list-decodable mean estimation at various scales.