Robust Hidden Markov Models Inference in the Presence of Label Noise

25. August 2014
CITEC 2.015


This talk shows how a statistical model can be made robust to expert errors. Indeed, the performance of traditional classification models can adversely be impacted by the presence of label noise in training observations. The pioneer work of Lawrence and Schölkopf tackled this issue in datasets with independent observations by incorporating a statistical noise model within the inference algorithm. In this talk, the specific case of label noise in non-independent, sequential observations is rather considered. In the first part of this talk, Hidden Markov models are reviewed. Learning algorithms are considered and their respective limitations are pointed out. Their use for automated electrocardiogram segmentation is discussed. Then, a label noise-tolerant expectation-maximisation algorithm is proposed. Experiments are carried on both healthy and pathological electrocardiogram signals with distinct types of additional artificial label noise. Results show that the proposed label noise-tolerant inference algorithm can improve the segmentation performances in the presence of label noise.