Santa Fe
Institute
  • Research
    • Themes
    • Projects
    • SFI Press
    • Researchers
    • Publications
    • Library
    • Sponsored Research
    • Fellowships
    • Miller Scholarships
  • News + Events
    • News
    • Newsletters
    • Podcasts
    • SFI in the Media
    • Media Center
    • Events
    • Community
    • Journalism Fellowship
  • Education
    • Programs
    • Projects
    • Alumni
    • Complexity Explorer
    • Education FAQ
    • Postdoctoral Research
    • Education Supporters
  • People
    • Researchers
    • Fractal Faculty
    • Staff
    • Miller Scholars
    • Trustees
    • Governance
    • Resident Artists
    • Research Supporters
  • Applied Complexity
    • Office
    • Applied Projects
    • ACtioN
    • Applied Fellows
    • Studios
    • Applied Events
    • Login
  • Give
    • Give Now
    • Ways to Give
    • Contact
  • About
    • About SFI
    • Engage
    • Complex Systems
    • FAQ
    • Campuses
    • Jobs
    • Contact
    • Library
    • Employee Portal

Science for a Complex World

Events

Here's what's happening

Give

You make SFI possible

Subscribe

Sign up for research news

Connect

Follow us on social media

© 2026 Santa Fe Institute. All rights reserved. This site is supported by the Miller Omega Program.

Home / News

Upending assumptions about learning, inspired by an AI phenomenon

Psychologists assume humans learn by simplifying their experiences, forgetting most things, and holding on only to key themes. Might it be possible, however, that we also learn through supercharged complexity? (image: Sabina Sloman with Gemini AI)
April 17, 2026

In the last decade, a phenomenon in artificial intelligence known as “double descent” has surprised researchers who study learning. The expectation was that the best AI models were neither too simple nor too complex. But certain models (like systems that classify images) actually learn best when the model grows extremely complex, far more than the data.

Over a century of psychological research assumed this was impossible: no system, animal or artificial, should be able to learn through such a method. Given constrained resources for storing information, systems should forget most of what they experience and only hold onto key themes.

Now that there’s proof of AI models successfully learning not through simplifying and forgetting, but through supercharged complexity, researchers are investigating whether humans can do something similar.

A recent paper in Brain and Behavioral Sciences aims to spark such investigations by proposing a new framework for human learning. Each year, the journal accepts a mere dozen or so manuscripts. Then it solicits up to 40 commentaries from outside experts, and publishes the robust scientific debate.

“We are challenging a long-standing assumption that says humans can’t learn complex representations and successfully generalize their knowledge,” says co-author and SFI Complexity Postdoctoral Fellow Marina Dubova, a cognitive scientist. “And we are proposing a framework that can be used to test if a certain human behavior actually demonstrates excess-capacity learning.”

The paper defines “capacity” as how many resources a learning system, like a human or AI model, devotes to representing its experiences. There are three kinds of capacity learning: constrained (where you don’t remember every experience), sufficient (where you remember every experience), and excess (where you remember every experience and supposedly irrelevant details about these experiences).

Imagine you want to learn what kind of food you like. In the traditional psychological “constrained” view, you forget the details of each meal you try, only remembering that you like burgers.

But one night, you have a bad burger. Excess-capacity learning means you memorize the exact details of the bad meal, and use these details to make predictions about future meals. Your predictions might depend on relevant details (the restaurant where you ate the bad burger) and details that may at first appear irrelevant (the day of the week).

Surprisingly, this excess complexity can help you make a better prediction about your next meal — while the day of the week alone may not directly affect burger quality, it could indicate which chef is working that day.

“Excess-capacity learning means that once you have memorized the exact details, you don’t need to discard the specifics to generalize. Instead, by retaining and even expanding on the details, you can arrive at a representation that generalizes well, sometimes even better than simplifying would. Cognitive science hasn’t fully explored this mode of learning,” says Dubova, who co-authored the paper with Sabina Sloman of the University of Manchester.

Psychological evidence already shows excess-capacity learning may exist in humans. Humans come up with more complex interpretations for phenomena than needed, and they often memorize redundant information.

Dubova and Sloman hope researchers who investigate intelligence and learning in the broadest sense — for example, biologists, philosophers, linguists, computer scientists, and cognitive scientists — will submit a commentary. Experts should register their interest by May 15, 2026.

Read the paper "Excess Capacity Learning" in Behavioral and Brain Sciences, April 17, 2026. DOI: 10.1017/S0140525X2610510X





Share
  • Sign Up For SFI News
News Media Contact

Santa Fe Institute

Office of Communications
news@santafe.edu
505-984-8800



  • Tags
  • SFI News Release


More SFI News

View All News

Why noise may be the key to understanding cell group patterns

Reinventing democracy before it breaks

Do deep learning models recognize 3D shapes in the same way humans do?

Upending assumptions about learning, inspired by an AI phenomenon

Looking at AGI through the lens of natural intelligence

A simple baseline for AI forecasting in machine learning

Constantino Tsallis to co-chair the 2027 Nobel Symposium on Statistical Mechanics

How novelty arrives: Review of “The Origins of the New”

Working group asks, what’s the benefit of a brain?

Measuring irreversibility in gene transcription

ACtioN Academy engages industry leaders on AI and complexity

Arguing for a complex adaptive power grid

Mark Newman Awarded 2026 SIAM John von Neumann Prize

Review: Nonesuch, by SFI Miller Scholar Francis Spufford

Laurent Hébert-Dufresne to receive Young Scientist Award

What does it mean to compute?

Reassessing the scientific method

SFI External Professor Santiago Elena elected to the American Academy of Microbiology

From cells to companies: Study shows how diversity scales within complex systems

SFI Press launches “The Economy as an Evolving Complex System IV”