Unifying Prediction, Idea Intervention, and Conditional Interpretations


[Submitted on 25 Jan 2024]

Obtain a PDF of the paper titled Power-Based mostly Idea Bottleneck Fashions: Unifying Prediction, Idea Intervention, and Conditional Interpretations, by Xinyue Xu and 4 different authors

Obtain PDF

Summary:Current strategies, corresponding to idea bottleneck fashions (CBMs), have been profitable in offering concept-based interpretations for black-box deep studying fashions. They usually work by predicting ideas given the enter after which predicting the ultimate class label given the anticipated ideas. Nevertheless, (1) they usually fail to seize the high-order, nonlinear interplay between ideas, e.g., correcting a predicted idea (e.g., “yellow breast”) doesn’t assist right extremely correlated ideas (e.g., “yellow stomach”), resulting in suboptimal closing accuracy; (2) they can’t naturally quantify the advanced conditional dependencies between totally different ideas and sophistication labels (e.g., for a picture with the category label “Kentucky Warbler” and an idea “black invoice”, what’s the chance that the mannequin accurately predicts one other idea “black crown”), subsequently failing to supply deeper perception into how a black-box mannequin works. In response to those limitations, we suggest Power-based Idea Bottleneck Fashions (ECBMs). Our ECBMs use a set of neural networks to outline the joint power of candidate (enter, idea, class) tuples. With such a unified interface, prediction, idea correction, and conditional dependency quantification are then represented as conditional chances, that are generated by composing totally different power features. Our ECBMs deal with each limitations of current CBMs, offering larger accuracy and richer idea interpretations. Empirical outcomes present that our method outperforms the state-of-the-art on real-world datasets.

Submission historical past

From: Xinyue Xu [view email]
Thu, 25 Jan 2024 12:46:37 UTC (5,233 KB)

Supply hyperlink


Please enter your comment!
Please enter your name here