Unveiling the Patterns: Measuring Human-AI Alignment in Model Behavior

Unveiling the Patterns: Measuring Human-AI Alignment in Model Behavior

Table of Contents:

1. Introduction
2. Evaluating Model Behavior 2.1 Selecting Inputs to Evaluate 2.2 The Problem of Subsampling
3. Introducing Shared Interest
4. Analyzing Model Behavior with Shared Interest 4.1 Quantifying Human Decision Making 4.2 Explaining Model Decision Making 4.3 Coverage Metrics: IoU Coverage, Saliency Coverage, and Ground Truth Coverage
5. Eight Common Cases of Model Behavior 5.1 Sufficient Subset Instances 5.2 Distractor Cases
6. Applying Shared Interest to Model Assisted Dermatology 6.1 Computer Vision Model for Melanoma Prediction 6.2 Analyzing Images with Shared Interest 6.2.1 Correctly Classified Images with High IoU Coverage 6.2.2 Context Dependent Images 6.2.3 Sufficient Context Cases
7. Querying Model Behavior with Shared Interest 7.1 Interactive Analysis with Ground Truth Regions 7.2 Uncovering Classes with High Alignment
8. Conclusion

Analyzing Model Behavior with Shared Interest

Shared Interest is a method that allows us to measure human-AI alignment and identify recurring Patterns in model behavior. In this article, we will explore how to evaluate model behavior and introduce the concept of Shared Interest. We will discuss how Shared Interest can be used to analyze model behavior, quantifying human decision making and explaining model decision making using coverage metrics. We will also explore the eight common cases of model behavior that can be uncovered using Shared Interest. Additionally, we will Apply Shared Interest to a real-world use case of model-assisted dermatology and discuss the implications of the analysis. Finally, we will Delve into querying model behavior using Shared Interest and conclude with the benefits of this approach in evaluating and deploying machine learning models.