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

Training a single sigmoidal neuron is hard.

Jirí Síma1

  • 1Institute of Computer Science, Academy of Sciences of the Czech Republic, PO Box 5, 18207 Prague 8, Czech Republic. sima@cs.cas.cz

Neural Computation
|November 16, 2002
PubMed
Summary
This summary is machine-generated.

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

On energy complexity of fully-connected layers.

Neural networks : the official journal of the International Neural Network Society·2024
Same author

Energy Complexity of Convolutional Neural Networks.

Neural computation·2024
Same author

Analog neuron hierarchy.

Neural networks : the official journal of the International Neural Network Society·2020
Same author

Subrecursive neural networks.

Neural networks : the official journal of the International Neural Network Society·2019
Same author

Energy complexity of recurrent neural networks.

Neural computation·2014
Same author

Sequential triangle strip generator based on Hopfield networks.

Neural computation·2009
Same journal

A Model-Free Reinforcement Learning Implementation of Decision Making Under Uncertainty by Sequential Sampling.

Neural computation·2026
Same journal

DROP: Distributional and Regular Optimism and Pessimism for Reinforcement Learning.

Neural computation·2026
Same journal

Hierarchical Active Inference Using Successor Representations.

Neural computation·2026
Same journal

W-Kernel and Its Principal Space for Frequentist Evaluation of Bayesian Estimators.

Neural computation·2026
Same journal

A Hidden Markov Model-Inspired Sequence Classification Method for Hyperdimensional Computing.

Neural computation·2026
Same journal

Sparse Graphical Modeling for Electrophysiological Phase-Based Connectivity Using Circular Statistics.

Neural computation·2026
See all related articles

Training even a single artificial neuron with a sigmoidal activation function is NP-hard. This finding implies that standard algorithms like backpropagation may not be efficient for simple neural network training.

Area of Science:

  • Artificial Intelligence
  • Machine Learning
  • Computational Complexity

Background:

  • Neural network training complexity is a significant area of research.
  • Previous studies have established hardness results for training feedforward neural networks.
  • Understanding the computational limits of training even simple network architectures is crucial.

Purpose of the Study:

  • To investigate the training hardness of the simplest neural network architecture: a single neuron with a sigmoidal activation function.
  • To mathematically prove the computational complexity of optimizing weights for this basic unit.
  • To assess the implications of these findings for existing learning algorithms.

Main Methods:

  • Survey of existing hardness results for feedforward neural network training.

Related Experiment Videos

  • Mathematical proof of NP-hardness for training a single sigmoidal neuron.
  • Analysis of minimizing quadratic training error and its average.

    • Main Results:

    • The problem of finding weights to minimize quadratic training error for a single sigmoidal neuron is NP-hard.
    • The specific error bounds for minimization are proven to be within (beta - alpha)(2) and 5(beta - alpha)(2)/ (12n).
    • The standard backpropagation algorithm is shown to be potentially inefficient even for this minimal case.

    Conclusions:

    • Training a single neuron with common sigmoidal activation functions is computationally intractable.
    • The inefficiency of backpropagation for single neurons has negative implications for constructive learning approaches.
    • Further research is needed to develop efficient learning algorithms for complex neural network architectures.