PROJECT 2

Deep learning and data geometry: A data-driven graph framework for explainable and trustworthy AI

Project Leader: Antonio Ortega

Website: https://viterbi.usc.edu/directory/faculty/Ortega/Antonio

Abstract: The overarching goal of this project is to develop insights into the geometry of input-output mappings learned by deep neural networks (DNN) and their impact on learning and generalization. We propose to (i) develop new neighborhood- and graph-based tools to capture the properties of feature space manifolds induced by DNNs and (ii) use them to analyze several key problems: (1) Trust: Can a model be considered trustworthy if its decisions and parameters are not robust to small changes in data? We propose to investigate the local geometric stability of data manifolds under perturbations in the data, algorithm, and model, including adversarial perturbations. (2) Explainability: Can we explain why a model reached a particular prediction? Is the decision biased? We propose to develop geometric explainability tools that characterize the local neighborhood of a query and the influence of training data on the decision. (3) Transfer: How do we determine if a model trained with one dataset or task can be used for another? We will leverage our proposed geometry-based methods to quantify dataset alignment and estimate the transfer learning capabilities of the model.