The Learning Circuit Nobody Had to Design
Ask how an artificial neural network learns and the answer begins with a decision somebody made. A person lays out the layers. A person writes the rule that adjusts each connection when the network gets something wrong. The brain is not built that way, and that is the awkward part of using one to explain the other: models of neural learning usually start with a circuit already arranged for the job, and say nothing about how biology would have arranged it.
A paper published on Oct. 9, 2026, in the Proceedings of the National Academy of Sciences takes the designer away and sees what is left. Qianli Liao, Liu Ziyin and Yulu Gan, the three co-first authors, worked with Mark T. Harnett and Tomaso Poggio at MIT's McGovern Institute. What they describe is a small unit of circuitry, a self-assembling motif, that nobody has to wire: it forms out of random connections on its own, and once formed it learns.
The argument around it matters as much as the mechanism. A network's value as a model of the brain, the team contends, should be judged by how much of itself it can build without help. "The brain doesn't have an external designer or architect," Liao says in the McGovern Institute's account of the work. "The brain has to organize itself. And we need to know how simple rules are going to make the brain organize."
The starting point is deliberately unhelpful. The model begins as several layers of neuron-like units joined at random, and every unit in a layer obeys the same local rule about when a link should strengthen or weaken. Nothing in that rule knows what the network is for. "They just interact with each other using that rule, and they develop some type of behavior," Liao says. Out of the interaction, the authors report, one arrangement keeps appearing: "The motif emerges from initially random connectivity under heterosynaptic plasticity rules," meaning rules where neighboring synapses change together.
What assembles is a pair of pathways rather than a single chain. One set of connections carries signals forward through the network; another carries them back the other way; a third set links the two. By the institute's account, both pathways reshape themselves at the same time, under the same local rules, until the backward one is sending the forward one signals it can actually learn from. In a hand-built network that feedback is the designer's contribution. Here it is the outcome.
The paper's title says where this is going: "How biological synapses self-assemble gradient learning." Stacked into a hierarchy, networks of these motifs settle into dynamics that approximate a generalized form of stochastic gradient descent, a method that shrinks errors. The authors say they prove the approximation rather than merely observe it. Most AI systems reach that method through backpropagation, which passes errors backward layer by layer. Neuroscientists mostly doubt the brain does anything of the sort. The team compared its self-organizing network with networks trained that way, and found it performed at a similar level.
The parity holds for the image-classification tasks reported, where a network sorts pictures into categories, and the paper claims nothing at the scale of the systems that write code or hold conversations. The authors claim a match, never an advantage. That is also the comfortable end of the problem: in the literature on biologically plausible learning rules, parity on small image sets has repeatedly given way to a gap on harder benchmarks.
The mechanism the whole thing rests on is where the paper is boldest. Heterosynaptic plasticity is real and well documented in cortical neurons, but the role that experiments have pinned on it is housekeeping: keeping connection strengths from running away, holding excitation and inhibition in balance. Promoting it to the thing that does the learning is this group's minority position, offered as the paper's contribution rather than as settled neuroscience.
And none of it has happened in a brain. Every result here is theory and simulation inside artificial networks; the paper proposes features of cortical wiring that experiments could look for, and nobody has looked. Poggio puts the conditional plainly: "If the synaptic motifs we propose are, in fact, used by the brain to do what backpropagation does for artificial neural networks, then we will have a strong connection at the level of fundamental mechanisms between artificial and natural networks—between machines and brains." Liao is careful in the same direction, saying a self-organizing network does not necessarily learn the way the brain does, while calling it a critical starting point.
What is new on Oct. 9, 2026, is the peer review rather than the idea. The group posted the self-assembly result as a preprint in December 2024, and the McGovern Institute published its own account on Oct. 5, 2026, four days before the journal version appeared. The code for the model is posted publicly, which is the part a skeptical lab can act on: run it, and watch what a different random start produces.
Liao does not oversell what the measure buys. "Self-organization is the first high-level framework," he says. "Once we are good with self-organization as a measure, then we can talk about next thing."