Fusing neural networks and graphical models in an active learning framework for image classification

Andrea Bertozzi
University of California, Los Angeles (UCLA)
Mathematics

This talk about similarity graphs for semi-supervised learning.
Data is represented by nodes on the graph and the graph weights measure pairwise similarity between individual pieces of data on the nodes.
Active learning involves providing labeled data (training data) as part of the algorithm using an acquisition function to select optimal graph nodes to label to provide the best accuracy. I will present recent successes with these methods for remote sensing applications using a variety of data sources including multimodal data. In complex imagery, unsupervised neural networks can provide superior graph weights for classification and in other cases, such as hyperspectral imagery pixel classification, a more standard cosine angle measure of the spectra works well. In all cases we find that these methods, when combined with active learning, can outperform state of art deep learning methods. I also will discuss some complications when these methods are applied to microscopy imagery.


Back to Workshop IV: Multi-Modal Imaging with Deep Learning and Modeling