Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

ARTIFICIAL NEURAL NETWORKS

An alternative to backpropagation through time

Nature Machine Intelligence volume 2, pages 155–156 (2020)Cite this article

Subjects

Recurrent networks can be trained using a generalization of backpropagation, called backpropagation through time, but a gap exists between the mathematics of this learning algorithm and biological plausibility. E-prop is a biologically inspired alternative that opens up possibilities for a new generation of online training algorithms for recurrent networks.

Backpropagation has already been generalized to recurrent neural networks based on exact mathematical minimization of the cost function, resulting in a method called back propagation through time (BPTT)3. BPTT considers the dependencies of the network’s output over previous time steps, unrolling the network activity backwards through time. This learning algorithm has been widely adopted by the machine learning community for training recurrent networks. However, it is unlikely that the brain performs BPTT, since it would need to store the full history of the network activity and consequently consume a large amount of energy.

This is a preview of subscription content, access via your institution

Access options

Rent or buy this article

Prices vary by article type

from$1.95

to$39.95

Learn more

Prices may be subject to local taxes which are calculated during checkout

Fig. 1: Recurrent neural networks are composed by loops that introduce temporal dependencies among nodes activities.

References

  1. Silver, D. et al. Nature 550, 354–359 (2017).

    Article  Google Scholar 

  2. Bellec, G. et al. Preprint at https://doi.org/10.1101/738385 (2019).

  3. Werbos, P. J. Proc. IEEE 78, 1550–1560 (1990).

    Article  Google Scholar 

  4. Clopath, C., Büsing, L., Vasilaki, E. & Gerstner, W. Nat. Neurosci. 13, 344–352 (2010).

    Article  Google Scholar 

  5. Williams, R. J. & Zipser, D. Neural Comput. 1, 270–280 (1989).

    Article  Google Scholar 

  6. Sutton, R. S. & Barto, A. G. Introduction to Reinforcement Learning Vol. 2 (MIT Press, 1998).

  7. Vasilaki, E., Frémaux, N., Urbanczik, R., Senn, W. & Gerstner, W. PLOS Comput. Biol. 5, e1000586 (2009).

    Article  Google Scholar 

Download references

Author information

Authors and Affiliations

  1. Department of Computer Science, University of Sheffield, Sheffield, UK

    Luca Manneschi & Eleni Vasilaki

Authors
  1. Luca Manneschi
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Eleni Vasilaki
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Eleni Vasilaki.

Ethics declarations

Competing interests

The authors declare no competing interests.

Rights and permissions

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Manneschi, L., Vasilaki, E. An alternative to backpropagation through time. Nat Mach Intell 2, 155–156 (2020). https://doi.org/10.1038/s42256-020-0162-9

Download citation

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s42256-020-0162-9

Search

Advanced search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing