InterActive Machine Learning

PI: Hsing-Kuo Kenneth Pao (National Taiwan University of Science and Technology), CoPI: Yuh-Jye Lee (National Chiao Tung University), Lin-Lin Chen (National Taiwan University of Science and Technology)

Champion: Omar Florez (Intel)

Project Objective: 

In the first year, we have designed the general optimization formulation for active learning that aims to work on the interactive learning scenarios. Some preliminary study have been finished for data that are from Internet of things and other domains. For the active learning strategy, we have developed a unify algorithm that considers uncertainty, diversity and representativity for data annotation. Between uncertainty and diversity, we gradually put more emphasis on the uncertainty as time goes by and it can give us the best performance with such a dynamically adjusted balance between uncertainty and diversity. More generally, we can discuss the scenario where more than a human (oracle) candidate can be consulted for the annotation tasks in a humans-machines integrated unit. Among different humans/oracles, we have various types such as between enthusiastic or reluctant, different availability, different data correctness, different costs, to name a few.

 

We continue brainstorming on the deployment of interactive machine learning techniques in with the design team. Some possible deployed scenarios includes the Smart homes, iFactory, intelligent transportation environment and others. Some current difficulties include how to collect data in an environment that fits to our interactive learning settings. Before it is possible to find such perfect scenarios, we also work on some well-known benchmark datasets that could be IoT or non-IoT types to justify the assumptions of the proposed methodology. (updated in Feb, 2017)