Descriptive Complexity for Neural Networks via Boolean Networks

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review

1 Citation (Scopus)
7 Downloads (Pure)

Abstract

We investigate the descriptive complexity of a class of neural networks with unrestricted topologies and piecewise polynomial activation functions. We consider the general scenario where the running time is unlimited and floating-point numbers are used for simulating reals. We characterize these neural networks with a rule-based logic for Boolean networks. In particular, we show that the sizes of the neural networks and the corresponding Boolean rule formulae are polynomially related. In fact, in the direction from Boolean rules to neural networks, the blow-up is only linear. We also analyze the delays in running times due to the translations. In the translation from neural networks to Boolean rules, the time delay is polylogarithmic in the neural network size and linear in time. In the converse translation, the time delay is linear in both factors. We also obtain translations between the rule-based logic for Boolean networks, the diamond-free fragment of modal substitution calculus and a class of recursive Boolean circuits where the number of input and output gates match.
Original languageEnglish
Title of host publication32nd EACSL Annual Conference on Computer Science Logic (CSL 2024)
EditorsAniello Murano, Alexandra Silva
Place of PublicationDagstuhl, Germany
PublisherSchloss Dagstuhl - Leibniz-Zentrum für Informatik
Pages9:1-9:22
Number of pages22
Volume288
ISBN (Electronic)978-3-95977-310-2
DOIs
Publication statusPublished - 27 Feb 2024
Publication typeA4 Article in conference proceedings
EventComputer Science Logic - Naples, Italy
Duration: 19 Feb 202423 Feb 2024

Publication series

NameLeibniz International Proceedings in Informatics (LIPIcs)
PublisherSchloss Dagstuhl -- Leibniz-Zentrum für Informatik
Volume288
ISSN (Print)1868-8969

Conference

ConferenceComputer Science Logic
Country/TerritoryItaly
CityNaples
Period19/02/2423/02/24

Publication forum classification

  • Publication forum level 1

Fingerprint

Dive into the research topics of 'Descriptive Complexity for Neural Networks via Boolean Networks'. Together they form a unique fingerprint.

Cite this