Restricted Research - Award List, Note/Discussion Page
Fiscal Year: 2023
2389 The University of Texas at San Antonio (144277)
Principal Investigator: Jha, Sumit
Total Amount of Contract, Award, or Gift (Annual before 2011): $ 160,000
Exceeds $250,000 (Is it flagged?): No
Start and End Dates: 8/15/22 - 8/14/24
Restricted Research: YES
Academic Discipline: SCIENCES
Department, Center, School, or Institute: Cos Computer Science
Title of Contract, Award, or Gift: Robust Explanations using Diverse Adversarially Trained Ensembles, Multi-Modal Contrastive Learning, and Attribution-based Confidence Metrics
Name of Granting or Contracting Agency/Entity:
University of Central Florida
CFDA: 0
Program Title: none
Note:
SAMs 1.1.1; In this research project, the UTSA team will perform the following tasks. Task 1 Ensembles of Diverse Neural Network and More Human-Interpretable Models using Mutual Information and Rényi Entropy as Diversity Metrics e.g., modular, stochastic, NAS, hyperparameters: A key challenge in creating ensembles of neural networks is to ensure their diversity. We will employ Rényi Entropy and greedy approximations of Shapley values to identify models whose addition to an ensemble enhances its diversity. We will investigate a new approach using stochastic generators such as DALL-E2 to create synthetic data sets that enable 100X faster adversarial training making them more readily usable for creating reliable explanations. Task 3 Multi-modal Shared Representations using Unsupervised/Contrastive Learning, such as text and images from the training data, counterfactuals, time series, trees, and rules/ODEs. Task 4 Attribution-based Confidence Metrics as Generalizations of Platt and Temperature Scaling.
Discussion: No discussion notes