Human-Machine Synergistic Learning

Pure data-driven machine learning approaches are usually black-boxes without the ability to improve their performance through interactions with humans.While an ideal learning algorithm should possess such an ability to reduce the heavy dependence over correctly labeled data and accurate problem modeling.To this end,we study a novel learning paradigm named human-machine synergistic learning,in order to make every stage of the learning system capable to interact with non-experts to improve their per-formance.In this work,we utilize crowdsourcing mechanism design,objective func-tion learning,and model reuse as examples to show how to deal with the limitations of human capability,meanwhile,make it better to satisfy the user needs.The main contributions are listed as follows:1.Crowdsourcing with unsure option.Considering the limitation of the labeling abil-ity of workers,in crowdsourcing,introducing the unsure option can improve label-ing accuracy significantly,while it can also lead to an increase in labeling cost.We conduct theoretical studies on the unsure mechanism for crowdsourcing.We show the sufficient condition under which the unsure mechanism can lead to significant performance improvement.We also design an effective online algorithm to realize the unsure mechanism based on our theoretical findings.2.Preference-based implicit objective learning.Considering the limitation of the mod-eling ability of users,in real machine learning applications,the system usually needs to be optimized under implicit objectives based on user satisfaction.These implicit objectives are hard to be analytically modeled and difficult to directly optimize.In this work,we proposed a hierarchical learning approach under the assumption that the preference-based feedback from the users is available,which can achieve effective learning on optimizing implicit learning objectives.3.Model Reuse from reusability indicator specifications.Considering the needs for target users to solve different tasks,in model reuse,designing model specifications is one of the essential challenges.Due to the unknown properties of target tasks,it is hard to guarantee the exact accuracy of the specifications learned from source data.We propose a model reuse algorithm based on reusability indicator specifications as well as an active rectification mechanism,which guarantees desired performance even when the specifications are significantly inaccurate.4.Model Reuse via Synergistic Training.Considering the need for model providers to protect their data privacy,in model reuse,there are strict restrictions on using the source model training data.We propose a novel model reuse approach based on the idea of synergistic training,which allows a group of source model providers to train a single selection model in cooperation without sacrificing their own data privacy,meanwhile accurate model reuse with reliable generalization abilities on both learning tasks and models is guaranteed.