Interpretable Maximal Discrepancies Metrics for Analyzing and Improving Generative Models
Abstract
The accuracy of machine learning systems for automatic object recognition depend on the degree to which training data represents the circumstances encountered during deployment. Training a system with representative samples enables it to effectively operate in a variety of environments without direct human intervention. As a system is deployed in evolving scenarios, it requires updating with new data to maintain performance. To maintain a representative sample, it is necessary to determine when new samples should be incorporated, as highly redundant instances or noisy data may unbalance the training data. This proposal concerns the analysis and implementation of algorithms for statistical divergence measures to compare two or more samples of data. In this project we will develop new methods to enable human interpretation of how two samples differ and algorithmic approaches to compensate for the differences, which may include the augmentation of the data by choosing from among other data collections or through the optimization of other machine learning or numerical software to create realistic data. A main focus of the research is developing precise and robust algorithms for identifying the specific instances or patterns that under or over-represented in one sample versus another. A second focus is studying how existing machine learning models can be leveraged to improve this identification. This will be done in the context of processing large amounts of images (derived from various modalities, such as acoustic imaging, specifically, synthetic aperture sonar), seeking to identify objects therein. The effort will focus on how statistical divergences operating on task-optimized features can be used for comparing training data sets and compensating for dataset shift, assessing generative models and augmenting their training, and estimating optimal parameter settings for physics-based simulations to enhance machine learning model performance through intelligent use of real and synthetic examples. The outcomes will be novel machine learning algorithms, their analysis and evaluation and comparison to the current stateof- the-art, detailed through publications and technical reports.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112300
Entities
People
- Austin J. Brockmeier
Organizations
- Office of Naval Research
- United States Navy
- University of Delaware