Public Abstract

DE-SC0023494: Robust Explanations using Diverse Adversarially Trained Ensembles, Multi-Modal Contrastive Learning, and Attribution-based Confidence Metrics

Award Status: Inactive

Institution: The University of Central Florida Board of Trustees, Orlando, FL
UEI: RD7MXJV7DKT9
PM: Fornari, Marco

Most Recent Award Date: 05/09/2025
Number of Support Periods: 2
PI: Ewetz, Rickard

Current Budget Period: 09/01/2023 - 08/31/2024
Current Project Period: 09/01/2022 - 08/31/2024

Public Abstract

Robust Explanations using Diverse Adversarially Trained Ensembles, Multi-Modal Contrastive Learning, and Attribution-based Confidence Metrics Public Project Summary Rickard Ewetz, University of Central Florida (Lead Principal Investigator) Sumit Kumar Jha, University of Texas at San Antonio (Co-Principal Investigator) Modern neural networks have surpassed human-level performance for many challenging tasks and are rapidly being deployed within real-world applications. The phenomenal performance of these networks comes from a large number of parameters and the non-linear interactions among them. This complex and high-dimensional dynamics makes it difficult for a human to understand and visualize why an AI system makes a particular decision. Robust methods for explaining AI decisions are needed to enhance the broad adoption of AI in scientific applications. While there has been a lot of work on explainable AI over the last decade, the emphasis often has been on the use of model gradients. We are not proposing a refinement to one of these gradient-based methods; instead, we are leveraging several new ideas, including symbolic reasoning, adversarially robust networks, synthesizing counterfactual inputs, multi-modal perceivers, self-supervised learning, and ensembles of diverse models. Our goal is to create robust multi-modal explanations using unsupervised learning for data including images, time series, text in the form of visualizations including super-pixels/text for images, and rules such as natural laws for time series. Robustness will be achieved by a combination of robust diverse ensembles of neural networks and faster adversarial training using stochastic generators. Attribution-based confidence metrics will be used to identify the envelope of the correct operation of neural networks. All work will be performed on state-of-the-art networks such as Transfomers and Perceivers. Our proposed effort will initiate a new wave of explainable AI. This will have a strong impact on many scientific disciplines where AI is being used for pursuing challenging tasks that ultimately require human trust, such as life sciences, deep reinforcement learning, physics, and cyber-physical systems. The project is also expected to broadly impact the scientific community by enabling the analysis of large datasets and models for scientific discovery.

Robust Explanations using Diverse Adversarially Trained Ensembles, Multi-Modal Contrastive Learning, and Attribution-based Confidence Metrics

Public Project Summary

Rickard Ewetz, University of Central Florida (Lead Principal Investigator)
Sumit Kumar Jha, University of Texas at San Antonio (Co-Principal Investigator)

Modern neural networks have surpassed human-level performance for many challenging tasks and are rapidly being deployed within real-world applications. The phenomenal performance of these networks comes from a large number of parameters and the non-linear interactions among them. This complex and high-dimensional dynamics makes it difficult for a human to understand and visualize why an AI system makes a particular decision. Robust methods for explaining AI decisions are needed to enhance the broad adoption of AI in scientific applications.

While there has been a lot of work on explainable AI over the last decade, the emphasis often has been on the use of model gradients. We are not proposing a refinement to one of these gradient-based methods; instead, we are leveraging several new ideas, including symbolic reasoning, adversarially robust networks, synthesizing counterfactual inputs, multi-modal perceivers, self-supervised learning, and ensembles of diverse models. Our goal is to create robust multi-modal explanations using unsupervised learning for data including images, time series, text in the form of visualizations including super-pixels/text for images, and rules such as natural laws for time series. Robustness will be achieved by a combination of robust diverse ensembles of neural networks and faster adversarial training using stochastic generators. Attribution-based confidence metrics will be used to identify the envelope of the correct operation of neural networks. All work will be performed on state-of-the-art networks such as Transfomers and Perceivers.

Our proposed effort will initiate a new wave of explainable AI. This will have a strong impact on many scientific disciplines where AI is being used for pursuing challenging tasks that ultimately require human trust, such as life sciences, deep reinforcement learning, physics, and cyber-physical systems. The project is also expected to broadly impact the scientific community by enabling the analysis of large datasets and models for scientific discovery.