Shared Interest: Measuring Human-AI Alignment to Identify Recurring Patterns in Model Behavior

要旨

Saliency methods --- techniques to identify the importance of input features on a model's output --- are a common step in understanding neural network behavior. However, interpreting saliency requires tedious manual inspection to identify and aggregate patterns in model behavior, resulting in ad hoc or cherry-picked analysis. To address these concerns, we present Shared Interest: metrics for comparing model reasoning (via saliency) to human reasoning (via ground truth annotations). By providing quantitative descriptors, Shared Interest enables ranking, sorting, and aggregating inputs, thereby facilitating large-scale systematic analysis of model behavior. We use Shared Interest to identify eight recurring patterns in model behavior, such as cases where contextual features or a subset of ground truth features are most important to the model. Working with representative real-world users, we show how Shared Interest can be used to decide if a model is trustworthy, uncover issues missed in manual analyses, and enable interactive probing.

受賞
Honorable Mention
著者
Angie Boggust
Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
Benjamin Hoover
IBM Research AI, Cambridge, Massachusetts, United States
Arvind Satyanarayan
MIT, Cambridge, Massachusetts, United States
Hendrik Strobelt
IBM Research AI, Cambridge, Massachusetts, United States
論文URL

https://dl.acm.org/doi/abs/10.1145/3491102.3501965

動画

会議: CHI 2022

The ACM CHI Conference on Human Factors in Computing Systems (https://chi2022.acm.org/)

セッション: Intelligent Systems, Human-AI Collaboration

383-385
5 件の発表
2022-05-04 01:15:00
2022-05-04 02:30:00