中央研究院 資訊科學研究所

How to recognize and model domain knowledge from data, and how to generate data representations comprehensive and helpful for experts and learners are challenging issues in combining artificial intelligence technology and expertise. Abundant information related to domain knowledge is usually interpreted and utilized in specific manners according to a small amount of observations, and the observation usually contains overlapped patterns and non-steady behaviors. Such tasks are hardly scalable because of the complexity of ground truth and the insufficiency of rigorously labeled datasets. Periodicity is arguably one of the most fundamental features in data streams. In particular, finding and extracting every oscillating component which has non-steady periods and oscillating patterns from multi-component data streams is still an open problem. In this talk, a newly proposed approach for mining non-steady periodicities is introduced. This approach includes the combined frequency and periodicity (CFP) method, the de-shaped synchrosqueezing transform (SST), together with an optimization approach for pattern extraction. The approach is reported useful and efficient in challenging problems including automatic music transcription for music information retrieval (MIR), motion artifact reduction in plethysmograph (PPG) signals, the separation of fetal electrocardiogram (ECG) signals, and other problems in future interdisciplinary technologies.