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.

  • NEWS AND VIEWS

Switching between tasks can cause AI to lose the ability to learn

Artificial neural networks become incapable of mastering new skills when they learn them one after the other. Researchers have only scratched the surface of why this phenomenon occurs — and how it can be fixed.
  1. Clare Lyle

    1. Clare Lyle is at Google DeepMind, Seattle, Washington, USA.

    View author publications

    You can also search for this author in PubMed  Google Scholar

  2. Razvan Pascanu

    1. Razvan Pascanu is at Google DeepMind, London, UK.

    View author publications

    You can also search for this author in PubMed  Google Scholar

Innovations in artificial intelligence (AI) are often driven by inspiration from biological systems. For example, the neural networks that form the basis of modern computer-vision systems share many similarities with the human visual system. But analogies between biological and artificial neural networks are often inexact, and nowhere is this gap more apparent than in continual learning, in which tasks are learnt sequentially through exposure to a changing stream of data, instead of through access to an entire fixed training data set. Writing in Nature, Dohare et al.1 report that neural networks frequently fail to learn continually, and suggest a solution that has already made waves in AI research.

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

Nature 632, 745-747 (2024)

doi: https://doi.org/10.1038/d41586-024-02525-z

References

  1. Dohare, S. et al. Nature 632, 768–774 (2024).

    Article  Google Scholar 

  2. McClelland, J. L., McNaughton, B. L. & O’Reilly, R. C. Psychol. Rev. 102, 419–457 (1995).

    Article  PubMed  Google Scholar 

  3. French, R. M. Trends Cogn. Sci. 3, 128–135 (1999).

    Article  PubMed  Google Scholar 

  4. Kirkpatrick, J. et al. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017).

    Article  PubMed  Google Scholar 

  5. Fedus, W. et al. Preprint at arXiv https://doi.org/10.48550/arXiv.2002.12499 (2020).

  6. Sokar, G., Agarwal, R., Castro, P. S. & Evci, U. in Proc. 40th Int. Conf. Machine Learning (ed. Krause, A. et al.) 32145–32168 (PMLR, 2023).

    Google Scholar 

  7. Lyle, C. et al. Proc. 40th Int. Conf. Machine Learning (ed. Krause, A. et al.) 23190–23211 (PMLR, 2023).

    Google Scholar 

Download references

Competing Interests

The authors declare no competing interests.

Subjects

Latest on:

Nature Careers

Jobs

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

Search

Advanced search

Quick links