Jiaqi Lv, Southeast University
Partial-label learning: Towards scalability and provability
To scale the acquisition of training data with limited overhead, a candidate label set from non-expert labelers is a popular surrogate for the exact expert-labeling for training instances. Learning a multi-class classifier from such inexact supervision is typically referred to as partial-label (PL) learning, which has various application domains, such as ecoinformatics, image annotation, web mining.
In the last two decades, most existing PL learning methods crafted learning objectives coupled to some specific optimization algorithms – there is neither theoretical understanding of the consistency nor practical scalability to big data. In this talk, I will introduce our recent work on PL learning. We proposed a progressive identification method that conducts the update of the model and identification of true labeling information seamlessly. Moreover, we proposed the first generation model of candidate label sets, and develop two novel partial label learning methods that are guaranteed to be provably consistent. All these methods are model-independent and loss-independent, and compatible with stochastic optimizer.
|Date||May 12, 2021 (Wed) 10:00 - 10:45|