Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

A machine learning method for extracting symbolic knowledge from recurrent neural networks.

A Vahed1, C W Omlin

  • 1Department of Computer Science, University of the Western Cape, Bellville, South Africa. avahed@uwc.ac.za

Neural Computation
|March 10, 2004
PubMed
Summary

This study introduces a new method for extracting knowledge from recurrent neural networks. It uses a symbolic learning algorithm to infer deterministic finite-state automata (DFAs) from network behavior, improving knowledge extraction fidelity.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Effects of home-based exercise training on gut microbiota and possible relations with cognitive function and metabolic health in postmenopausal women with type 2 diabetes mellitus: a randomized control trial.

Diabetes research and clinical practice·2025
Same author

Stable encoding of large finite-state automata in recurrent neural networks with sigmoid discriminants.

Neural computation·1996
Same author

Pruning recurrent neural networks for improved generalization performance.

IEEE transactions on neural networks·1994
See all related articles

Area of Science:

  • Artificial Intelligence
  • Machine Learning
  • Computational Neuroscience

Background:

  • Neural networks are often 'black boxes,' making it difficult to interpret the knowledge within their weights.
  • Existing methods for extracting symbolic knowledge from neural networks, particularly recurrent neural networks (RNNs), face challenges with computational complexity and potential loss of fidelity.
  • Current approaches often rely on state clustering assumptions, which can be computationally intensive and may require limiting network complexity or exploration.

Purpose of the Study:

  • To develop a novel method for extracting symbolic knowledge from trained recurrent neural networks (RNNs).
  • To infer deterministic finite-state automata (DFAs) from RNNs without relying on state clustering hypotheses.
  • To improve the fidelity of extracted knowledge compared to existing methods.

Related Experiment Videos

Main Methods:

  • The study proposes a polynomial time, symbolic learning algorithm.
  • This algorithm infers DFAs directly from the observed input-output behavior of a trained RNN.
  • It avoids the computational complexity and limitations associated with cluster analysis of network states.

Main Results:

  • The proposed method offers a computationally efficient approach to knowledge extraction.
  • It has the potential to increase the fidelity of the extracted symbolic knowledge.
  • The algorithm infers DFAs solely from input-output observations, bypassing the need for internal state analysis.

Conclusions:

  • The novel symbolic learning algorithm provides a more faithful method for extracting knowledge from RNNs.
  • This approach overcomes the limitations of previous cluster-based methods.
  • It enhances the interpretability of recurrent neural networks by accurately representing their learned behavior as DFAs.