Image Quality and Task Complexity for Machine Learning

This STTR Phase I project proposes propose a novel framework for image quality assessment for generalized evaluation to enable efficient and model-agnostic assessment of machine learning models based on objective image quality. This is a generalized framework for ML performance prediction across various models and tasks. Our approach assesses the intermediate representations of input patterns and their impact on accurate predictions of task-specific outputs. By characterizing these representations through their dependencies on the input and output, the proposed framework quantifies mutual information, which is invariant to invertible transformations, making it applicable across various machine learning models regardless of their architecture. The Phase I proposed solutions and resulting system will be built upon the extensive experience of the Novateur team in information theory, image quality assessment, deep convolutional neural networks for object detection and classification in ISR imagery, low-shot object detection, domain adaptation, metric learning, and performance assessment of machine learning approaches.