论文笔记 Interpreting Black-Box Classifiers Using Instance-Level Visual Explanations

Subject: Interactive Model Analysis

Target: Verify the performance of a model

Existing methods: statistical methods, in an aggregated fashion (e.g. accuracy)

Related work:

White box approach: Aiming at visualizing the internal structures of the models

　　Logistic Regression: transparent weighting of the features

Black box approach

Models comparison:

　　ModelTracker

MLCube Explorer: data cube analysis type

Contribution: a workflow and an interface

Novelty

Focus on input/output behaviour of a model (model agnostic)

Locally and globally, decisions and feature importance

Workflow:



Core of the explanation algorithm: Removing features from a vector until the predicted label changes.

User Interface of Rivelo



Limitations: works with binary classifiers and binary features

Useful Quotes: DARPA XAI program: “the effectiveness of these systems is limited by the machines current inability to explain their decisions and actions to human users [. . .] it is essential to understand, appropriately trust, and effectively manage an emerging generation of artificially intelligent machine partners"

Reference:

[1] Tamagnini, Paolo, et al. "Interpreting Black-Box Classifiers Using Instance-Level Visual Explanations." (2017).