ub.xmlui.mirage2.page-structure.muninLogoub.xmlui.mirage2.page-structure.openResearchArchiveLogo
    • EnglishEnglish
    • norsknorsk
  • Velg spraakEnglish 
    • EnglishEnglish
    • norsknorsk
  • Administration/UB
View Item 
  •   Home
  • Fakultet for naturvitenskap og teknologi
  • Institutt for fysikk og teknologi
  • Artikler, rapporter og annet (fysikk og teknologi)
  • View Item
  •   Home
  • Fakultet for naturvitenskap og teknologi
  • Institutt for fysikk og teknologi
  • Artikler, rapporter og annet (fysikk og teknologi)
  • View Item
JavaScript is disabled for your browser. Some features of this site may not work without it.

The computational complexity of understanding binary classifier decisions

Permanent link
https://hdl.handle.net/10037/24133
DOI
https://doi.org/10.1613/JAIR.1.12359
Thumbnail
View/Open
article.pdf (712.5Kb)
Published version (PDF)
Date
2021-01-21
Type
Journal article
Tidsskriftartikkel
Peer reviewed

Author
Wäldchen, Stephan; Macdonald, Jan; Hauch, Sascha; Kutyniok, Gitta Astrid Hildegard
Abstract
For a d-ary Boolean function Φ: {0, 1}d → {0, 1} and an assignment to its variables x = (x1, x2, . . . , xd) we consider the problem of finding those subsets of the variables that are sufficient to determine the function value with a given probability δ. This is motivated by the task of interpreting predictions of binary classifiers described as Boolean circuits, which can be seen as special cases of neural networks. We show that the problem of deciding whether such subsets of relevant variables of limited size k ≤ d exist is complete for the complexity class NPPP and thus, generally, unfeasible to solve. We then introduce a variant, in which it suffices to check whether a subset determines the function value with probability at least δ or at most δ − γ for 0 < γ < δ. This promise of a probability gap reduces the complexity to the class NPBPP. Finally, we show that finding the minimal set of relevant variables cannot be reasonably approximated, i.e. with an approximation factor d1−α for α > 0, by a polynomial time algorithm unless P = NP. This holds even with the promise of a probability gap.
Publisher
AI Access Foundation
Citation
Wäldchen, Macdonald, Hauch, Kutyniok. The computational complexity of understanding binary classifier decisions. The journal of artificial intelligence research. 2021;70:351-3987
Metadata
Show full item record
Collections
  • Artikler, rapporter og annet (fysikk og teknologi) [1057]
©2021 AI Access Foundation. All rights reserved.

Browse

Browse all of MuninCommunities & CollectionsAuthor listTitlesBy Issue DateBrowse this CollectionAuthor listTitlesBy Issue Date
Login

Statistics

View Usage Statistics
UiT

Munin is powered by DSpace

UiT The Arctic University of Norway
The University Library
uit.no/ub - munin@ub.uit.no

Accessibility statement (Norwegian only)