Robustness at Inference: Towards Explainability, Uncertainty, and Intervenability

Event: CVPR 2024 Tutorial · Duration: 253 min · ▶ Watch on YouTube

Abstract

This tutorial explores robustness at inference in neural networks, focusing on explainability, uncertainty, and intervenability. It delves into various methods for quantifying and reducing uncertainty, evaluating explanations, and understanding the impact of interventions. The presentation covers theoretical foundations, practical applications, and case studies to illustrate the challenges and advancements in building robust and interpretable AI systems.

Speakers

  • Ghassan AlRegib — Georgia Institute of Technology
  • Mohit Prabhushankar — Georgia Institute of Technology

Talks (4)

  • 00:00:00 — Mohit Prabhushankar: Robustness at Inference: Towards Explainability, Uncertainty, and Intervenability
    • Introduction to the tutorial, covering the agenda, logistics, and the core concept of robustness at inference.
  • 01:11:53 — Mohit Prabhushankar: Uncertainty: What is Uncertainty?
    • Explores the definition of uncertainty in AI models, distinguishing between aleatoric and epistemic uncertainty, and how it’s quantified using ensembles and Monte Carlo dropout.
  • 02:39:39 — Mohit Prabhushankar: Uncertainty: Gradients as Single pass Uncertainty Quantification
    • Discusses using gradients to characterize novel data uncertainty at inference, focusing on backpropagating confounding labels for out-of-distribution detection.
  • 03:05:55 — Mohit Prabhushankar: Intervenability Frameworks: Dangers of Incomplete Interventions: SHAPE Explanations
    • Introduces intervenability frameworks, highlighting the dangers of incomplete interventions and how SHAPE explanations can yield unexpected results.

Methods / Models / Datasets Mentioned

  • WCB
  • GPT-3
  • GPT-4
  • Alexnet
  • VGG-16
  • ResNet-18
  • ResNet-34
  • ResNet-50
  • ResNet-101
  • DenseNet
  • SWIN Transformer
  • GradCAM
  • GradCAM++
  • RISE
  • SHAPE
  • PointPrompt
  • SAM
  • Mahalanobis
  • ODIN
  • LRP
  • SHAP

Topics

Robustness at Inference · Explainability · Uncertainty Quantification · Intervenability · Neural Networks · Gradient Analysis · Monte Carlo Dropout · Out-of-Distribution Detection · Causal Analysis · Evaluation Metrics

Notes

Open for commentary — connections to other work, critiques, follow-up reading.