Title: A Multi-Kernel Active Learning Framework
Presenting Author: James C. Tilton
Organization: NASA GSFC
Co-Author(s): Yuhang Zhang/U of Houston, H. Lexie Yang/Purdue, Edoardo Pasolli/Purdue, Zhou Zhang/Purdue,
Melba M. Crawford/Purdue, Saurabh Prasad/U of Houston

Abstract:
Classification of high dimensional data acquired by single or multiple sensors is an increasingly relevant problem for mapping applications, but requires extensive labeled data for training algorithms and is computationally intensive. This project leverages multiple kernel based machine learning for exploiting disparate features and advances in the Hierarchical Segmentation (HSeg) method in an Active Learning (AL) framework to facilitate construction of an effective training pool. Multiple-kernel based machine learning effectively exploits disparate features from single or multiple sources, while reducing the computational burden of the AL implementation. Hyperparameter tuning required by mixture-of-kernel approaches for model selection and kernel combination functions may require tuning at each AL learning step, which is potentially very time-consuming. In this project, the proposed multiple kernel active learning algorithm promises enhanced classification, improved AL performance, and a mechanism for automatic selection of kernel weights in the mixture-of-kernels. We have demonstrated the utility of the proposed framework with results for both feature fusion and sensor fusion tasks. Additionally, we have advanced the capability of HSeg, providing new capability for reducing over-segmentation and adaptively pruning the segmentation tree. HSeg has also been integrated within the proposed framework to effectively integrate spatial contextual information for the AL tasks while reducing the computational overhead.