Randomly wired units in an MIT network assemble their own learning machinery
MIT researchers report in PNAS that randomly connected units can build their own learning machinery from simple rules for when connections change. The authors propose self-organization as the test of whether a network is biologically plausible.
The Scientist · Science desk
What happened
- The network starts as multiple layers of randomly interconnected neuron-like units, and every unit within a layer follows the same rules for adjusting its connections.
- Among those rules is 'neurons that fire together wire together': when two connected neurons activate at the same time, the link between them strengthens.
- In the team's model, learning runs on motifs that assemble themselves under the simple rules and then act as building blocks for functional circuits.
- Trained this way, the network learned to recognize patterns and classify images.
- Against networks designed to learn with back-propagation, the self-organizing network performed at a similar level, according to MIT.
Why it matters
- decision Neuroscientists choosing networks as models get a different ranking criterion: the authors say a network's value should be judged by how far it organizes itself, meaning both its connections and the machinery that adjusts them.
- contradiction Back-propagation is the usual implementation of gradient descent and is thought unlikely in the brain, yet the network's motifs still reorganize with a strategy similar to gradient descent. Error-driven adjustment stays in the loop, so the plausibility claim concerns where the machinery comes from.
- constraint The thing this doesn't tell you is whether the brain learns this way. Matching back-propagation networks on image classification shows the self-built machinery works; a model that works can still differ from the brain.
The design answers a problem with networks whose architecture a developer supplies. The authors caution that networks built that way may not offer much insight into biological learning [14]. Liao, one of the paper's authors [16], said the brain has no such designer: "The brain doesn't have an external designer or architect," Liao said. "The brain has to organize itself. And we need to know how simple rules are going to make the brain organize." [11]
The comparison with back-propagation networks is the closest thing to a control in MIT's account. It addresses the cheapest objection, that a network which builds its own machinery must learn worse. How close the match is cannot be read from the summary, which does not list the image datasets or the scores, or say whether results hold across different random starting connections.
We think self-organization is a sensible filter for models. A network that needs a designer to supply its learning machinery cannot show how simple rules produce that machinery.
Liao ties the question to brain-computer interfaces, devices meant to give people back the ability to communicate after the brain loses its capacity for language, or to let paralyzed patients operate neuroprosthetic limbs [12]. Designing a circuit that performs well is not enough there. "None of these can work without knowing the rules behind neuron-neuron interactions," he said. "That's not achievable by just designing a circuit that works very well." [13]
What to watch
- Whether the full PNAS paper reports the image datasets, the scores, and the spread across random starting connections behind the back-propagation comparison.
- Whether the approach holds up when different layers carry different rules, since the rules may differ across brain regions or change with maturity [c17].
- Whether the learned rules can be checked against measured neuron-neuron interactions, which Liao says brain-computer interface work depends on.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+10
- Incentives
- Insufficient
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The paper was published Oct. 9 in the journal PNAS by scientists at MIT's McGovern Institute for Brain Research and K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center.
- [2]
The authors argue that a crucial metric for the biological plausibility of an artificial neural network's approach to learning is the network's ability to organize itself, without the guiding hand of a developer.
- [3]
With simple instructions about how and when connections should change, a group of randomly connected units can reshape itself into an effective information-processing network, learning from the data it is trained on.
- [4]
The network's structure begins with multiple layers of randomly interconnected neuron-like units, and within each layer every unit is subject to the same rules about how connections should be adjusted.
- [5]
The network implements simple rules thought to rewire brain circuits, such as 'neurons that fire together wire together': when two connected neurons are activated at the same time, the connection between them strengthens.
- [6]
In the team's model, learning is enabled by self-assembling motifs that emerge according to simple rules and then serve as building blocks for functional circuits.
- [7]
As the network learned, its motifs reorganized using an optimization strategy similar to gradient descent, in which connections are gradually adjusted in response to errors to improve performance.
- [8]
Most modern AI implements gradient descent through some form of back-propagation, which neuroscientists say is unlikely to be used by the brain.
- [9]
The team compared its self-organizing network to networks designed to learn with back-propagation and found that it performed at a similar level.
- [10]
The team showed that a self-organizing network governed by such rules can learn to recognize patterns and classify images.
- [11]
"The brain doesn't have an external designer or architect," Liao said. "The brain has to organize itself. And we need to know how simple rules are going to make the brain organize."
- [12]
Liao points to the development of brain-computer interfaces, which aim to restore communication when the brain's language abilities are lost or let paralyzed patients control neuroprosthetic limbs, as one potential application.
- [13]
"None of these can work without knowing the rules behind neuron-neuron interactions," Liao said. "That's not achievable by just designing a circuit that works very well."
- [14]
Liao and his colleagues caution that networks designed with a developer-built architecture may not offer much insight into biological learning.
- [15]
An artificial neural network's value to neuroscience should be judged by the extent to which it can organize itself, not just its connections but also the machinery that adjusts those connections.
- [16]
Working with Poggio and Hartnett, Liao, Ziyin and Gan set out to develop and validate a self-organizing computational network capable of learning.
- [17]
The rules that reshape neural connections as an animal learns may be different in different parts of the brain, or may change as the brain becomes more mature.
Sources
1 independent publisher whose own reporting we read for this story.
- news.mit.eduHow artificial neural networks build their own learning machinery
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.