Quantum Information Potential Fields: ANovelUncertainty Quantification Framework for Machine Learning
Abstract
In this grant, we will formulate uncertainty in a novel way inspired by information theoretic learning, more specifically in wave decompositions of local kernel-based realizations of distributions learnt by the model and the chosen parameters. The decompositions will be performed via the Schroedinger’s equation and Hermite decompositions of the model output and its intermediate representations of the data; decompositions can be easily and quickly rederived for each new input sample. Such an uncertainty measure is non-intrusive to the training process of point-wise prediction models, as it relies solely on extracting details from the internal dependencies of the trained model with respect to its output. Our preliminary results show that the method can be used efficiently in online applications. During the grant, we will advance our formulation of quantum-based uncertainty, with the goal of providing a methodology for parameter and also output uncertainty that can be implemented in an online mode. We will demonstrate that it can be applied to either spatial or non-stationary temporal and spatio-temporal data, along with either shallow or deep models without the need to directly estimate probability densities. It thus will be highly general compared to existing uncertainty quantification methods, especially those for neural networks. We plan to compare our measure with others in the literature to highlight its robustness and computational efficiency. We also will apply this measure to automated target analysis problems, such as target detection and recognition, seafloor semantic segmentation, along with scene change summarization. We will emphasize its use for both training and assessing models for under-utilized imaging sonar modalities, such as circular-scan synthetic aperture sonar and forward-looking sonar.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112345
Entities
People
- José Príncipe
Organizations
- Office of Naval Research
- United States Navy
- University of Florida