|
Abstract:
|
Inference with fusion data of disparate sources, which usually live in high dimensional space, is problematic given the curse of dimensionality, scaling, and over fitting concerns. Often, feature selection is performed for data reduction. The conditionality principle states that conditioning inference on ancillary statistics can improve statistical power. Iterative denoising trees (IDT) have been shown beneficial for driving data discovery. IDT are constructed through hierarchical conditioning without regard to labels and subsequent classification. Here, we explore using IDT to identify image content by combining statistics derived from image and caption features. When compared against Fisher's linear discriminant, IDT enable more accurate classification with limited features. In summary, IDT is a promising method to exploit conditioning for deriving new classifiers.
|
