Modern AI is trained with backpropagation, an algorithm that most neuroscientists think brains do not use. A paper published in the Proceedings of the National Academy of Sciences on October 9 asks a different question: could a brain-like network assemble its own learning machinery, starting from random connections and simple local rules, and still end up learning about as well as a standard network? The study is peer reviewed, and it is a theoretical and computational result rather than an experiment on living tissue.

The problem the paper tackles

Backpropagation works by computing how much each connection in a network contributed to an error and adjusting it accordingly. That requires carefully designed machinery: error signals travelling backwards through exactly the right connections, in a precise sequence. Neuroscientists have long doubted that biology does this. Existing proposals for brain-compatible learning rules usually take a circuit that is already arranged correctly and show how it could learn. The authors point out that this skips a step: it explains how a given circuit learns, but not how biology could assemble that circuit in the first place.

As co-first author Qianli Liao put it in MIT’s coverage, “The brain doesn’t have an external designer or architect. The brain has to organize itself.”

What the authors propose

The paper formulates what the authors call Self-Assembling Learning, the study of how learning systems can emerge from lower-level interactions. As one example mechanism, they introduce the Self-Assembling Motif, or SAM. According to the abstract, SAM self-assembles at two scales. First, the motif emerges from initially random connectivity under heterosynaptic plasticity rules, meaning rules in which the strength of a connection can change depending on the activity of neighboring connections, not only on the two neurons it links. Second, networks of SAMs, composed hierarchically, self-organize into dynamics that the authors say “provably approximate a generalized form of stochastic gradient descent,” matching backpropagation-level performance.

In plain terms, you start with layers of randomly wired neuron-like units. Every unit follows the same simple rules about how connections should strengthen or weaken. Out of those interactions, small repeating structures form, and those structures act as building blocks that, together, carry out something equivalent to gradient descent. MIT’s write-up notes that the rules include the familiar idea that neurons that fire together wire together, and that the team showed such a network can learn to recognize patterns and classify images. Compared against networks designed to learn with standard backpropagation-style training, the self-organizing network performed at a similar level, according to MIT.

Why this is interesting for AI as well as neuroscience

For neuroscience, the claim is that learning does not have to be specified in detail by genes or designers. It could emerge from local interactions, and, in the authors’ words, “to a far greater extent than previously thought.” Tomaso Poggio, one of the senior authors, told MIT that finding a biological parallel to backpropagation could trigger significant progress in both fields: if the proposed motifs are what the brain uses to do what backpropagation does for artificial networks, “then we will have a strong connection at the level of fundamental mechanisms between artificial and natural networks.”

For AI, the interest is more speculative. Current training is powerful but depends on global bookkeeping across the whole network, which is expensive on today’s hardware. Learning schemes built from local rules might one day suit different kinds of hardware, or help explain why trained networks and brains sometimes represent information similarly. Nothing in the paper suggests it will replace backpropagation in practical systems soon. Our explainers on gradient descent and machine learning basics cover the background.

How to read the result

A few cautions are worth keeping in mind. First, matching backpropagation “in performance” was demonstrated on the benchmark tasks the authors used, which the coverage describes as pattern recognition and image classification; this is not the same as matching it at the scale of today’s large language models. Second, a mechanism that is plausible in simulation is not proof that brains use it. Liao himself says a self-organizing network “does not necessarily learn the same way the brain does,” and calls self-organization “the first high-level framework,” a starting point for judging biological plausibility. Third, the authors argue that a good test of a brain-inspired model should be whether it can organize itself, not only whether its final circuit works, which is a useful change in how such models are judged.

What to watch next

  • Experimental neuroscience tests that look for motif-like structures and heterosynaptic rules in real tissue.
  • Whether the approach scales to deeper networks and harder tasks than image classification.
  • Follow-up work on whether biologically plausible learning rules can run efficiently on neuromorphic or other non-standard hardware.
  • Applications the researchers mention, such as understanding how learning rules change with aging or disease and designing brain-computer interfaces.
  • Critiques from researchers who favor other proposals for brain-compatible backpropagation.

Key takeaways

  • The paper is peer reviewed in PNAS (volume 123, issue 41) and comes from MIT’s McGovern Institute and collaborators.
  • It argues that brain-like learning circuits can self-assemble from random connectivity under simple local rules.
  • The resulting networks provably approximate a form of stochastic gradient descent and match backpropagation-level performance on the tested tasks.
  • It does not show that the brain actually learns this way.
  • Scaling to modern large models is untested.

Frequently asked questions

Does this mean backpropagation is wrong?

No. Backpropagation remains the standard way to train artificial networks. The paper explores whether biology could achieve something equivalent without a designer.

What is heterosynaptic plasticity?

A change in the strength of one connection that depends on the activity of other nearby connections, instead of only on the two neurons directly linked.

Will this change how AI models are trained?

Not in the near term. The immediate relevance is to neuroscience, with possible long-term ideas for new learning hardware.