In a paper published on 9 October in PNAS, researchers at MIT’s McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience Center propose a metric for judging whether an artificial neural network tells you anything about the brain: can it organise itself, without a designer?
Then they built one that does.
The argument first
Artificial networks and brain circuits both reorganise as they learn, adjusting connection strengths to improve performance. That resemblance is why computational models are used in neuroscience at all. But the resemblance hides a difference that the paper’s authors think is decisive.
In a conventional network, engineers design the rules about how connections change as training data arrives and errors come back. The rules are the researcher’s contribution. As co-first author Qianli Liao — a K. Lisa Yang ICoN Center Postdoctoral Fellow in the labs of Tomaso Poggio and Mark Harnett — puts it, that approach is powerful for getting AI to do a job, and may offer little insight into biological learning.
“The brain doesn’t have an external designer or architect,” Liao says. “The brain has to organize itself. And we need to know how simple rules are going to make the brain organize.”
So the proposed criterion is not accuracy, and not even biological plausibility of the architecture. It is whether the network can assemble not just its connections, but the machinery that adjusts those connections. A model where the learning algorithm is handed down from outside has quietly imported the thing you were trying to explain.
What they built
Liao, with Liu Ziyin (MIT research scientist, co-first author) and graduate student Yulu Gan, working with Poggio and Harnett, developed and validated a self-organising network that learns.
The structure starts as multiple layers of randomly interconnected neuron-like units. Within each layer, every unit is subject to the same rule about how connections should be adjusted. Nothing is placed by hand. “They just interact with each other using that rule, and they develop some type of behavior,” Liao says.
The rules implemented are the ones long thought to rewire brain circuits during learning — most famously the Hebbian principle that neurons that fire together wire together: when two connected neurons activate simultaneously, the connection between them strengthens.
From those local rules, self-assembling motifs emerge — small recurring patterns of connectivity that then serve as building blocks for functional circuits. The network learned to recognise patterns and classify images.
The result that makes it interesting
As the network learned, those motifs reorganised themselves using an optimisation strategy similar to gradient descent — the method at the centre of essentially all modern machine learning, in which connections are nudged in response to errors.
This matters because of a long-standing impasse. Modern AI implements gradient descent through back-propagation, and neuroscientists have argued for decades that back-propagation is unlikely to be what the brain does: it requires error signals to travel backwards through the exact forward pathway, with precise weight symmetry that biological synapses have no obvious way to maintain.
The team compared its self-organising network against networks trained conventionally and found it performed at a similar level. The paper’s title states the claim directly: How Biological Synapses Self-Assemble Gradient Learning.
If gradient-like learning can emerge from local rules rather than being imposed by a global algorithm, the objection to back-propagation loses some of its force — not because the brain does back-propagation, but because the brain might not need to in order to get gradient learning’s benefits.
Poggio’s framing is that finding a biological parallel to back-propagation could trigger significant progress in both fields at once.
Why it is not only a neuroscience result
Liao points to brain-computer interfaces as the application that depends on this kind of knowledge — restoring communication when language abilities are lost, or letting paralysed patients control neuroprosthetic limbs. “None of these can work without knowing the rules behind neuron-neuron interactions,” he says. “That’s not achievable by just designing a circuit that works very well.”
That distinction is the whole paper in one line. A circuit that works is an engineering result. A circuit whose organising rules you understand is a result you can intervene on — and intervention is what a prosthesis is.
Liao is careful about the scope: “Self-organization is the first high-level framework. Once we are good with self-organization as a measure, then we can talk about next thing.” The rules that reshape connections may differ between brain regions, and may change as a brain matures — none of which this work resolves.
The group’s primary goal is understanding brains. But Liao notes the knowledge could also improve machine learning, letting engineers build networks that perform better, or in more human-like ways.
For readers here
The transferable idea is the methodological one, and it generalises past neuroscience. A system that behaves correctly because a designer specified the behaviour explains less than one that arrives at the behaviour from simple local rules. Anyone who has written a generative system knows the version of this: hand-tuned output can be made to look like anything, and a result that emerges from a small rule set is the one that teaches you something — and the one that keeps surprising you.
That is the connection to generative practice — L-systems, cellular automata, flocking — where the whole discipline is choosing rules and letting structure arise. This paper applies the same standard to learning itself.
Sources
- How artificial neural networks build their own learning machinery — MIT News
- How Biological Synapses Self-Assemble Gradient Learning — PNAS, 10.1073/pnas.2606245123
- McGovern Institute for Brain Research
- K. Lisa Yang Integrative Computational Neuroscience Center