Saved in:
Bibliographic Details
Main Authors: Wu, Min, Li, Xiaofu, Wu, Haoze, Barrett, Clark
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2409.03060
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908655008350208
author Wu, Min
Li, Xiaofu
Wu, Haoze
Barrett, Clark
author_facet Wu, Min
Li, Xiaofu
Wu, Haoze
Barrett, Clark
contents Building on VeriX (Verified eXplainability, arXiv:2212.01051), a system for producing optimal verified explanations for machine learning models, we present VeriX+, which significantly improves both the size and the generation time of formal explanations. We introduce a bound propagation-based sensitivity technique to improve the size, and a binary search-based traversal with confidence ranking for improving time -- the two techniques are orthogonal and can be used independently or together. We also show how to adapt the QuickXplain algorithm to our setting to provide a trade-off between size and time. Experimental evaluations on standard benchmarks demonstrate significant improvements on both metrics, e.g., a size reduction of $38\%$ on the GTSRB dataset and a time reduction of $90\%$ on MNIST. We demonstrate that our approach is scalable to transformers and real-world scenarios such as autonomous aircraft taxiing and sentiment analysis. We conclude by showcasing several novel applications of formal explanations.
format Preprint
id arxiv_https___arxiv_org_abs_2409_03060
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Efficiently Computing Compact Formal Explanations
Wu, Min
Li, Xiaofu
Wu, Haoze
Barrett, Clark
Machine Learning
Artificial Intelligence
Building on VeriX (Verified eXplainability, arXiv:2212.01051), a system for producing optimal verified explanations for machine learning models, we present VeriX+, which significantly improves both the size and the generation time of formal explanations. We introduce a bound propagation-based sensitivity technique to improve the size, and a binary search-based traversal with confidence ranking for improving time -- the two techniques are orthogonal and can be used independently or together. We also show how to adapt the QuickXplain algorithm to our setting to provide a trade-off between size and time. Experimental evaluations on standard benchmarks demonstrate significant improvements on both metrics, e.g., a size reduction of $38\%$ on the GTSRB dataset and a time reduction of $90\%$ on MNIST. We demonstrate that our approach is scalable to transformers and real-world scenarios such as autonomous aircraft taxiing and sentiment analysis. We conclude by showcasing several novel applications of formal explanations.
title Efficiently Computing Compact Formal Explanations
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2409.03060