Artificial neural networks are usually shaped by engineers: designers decide how connection strengths change as the network learns from data and feedback. A new study from MIT argues that this designer-driven approach, however powerful, may tell us little about how the brain actually learns — and demonstrates an alternative in which networks organize themselves.
In a paper published October 9 in the Proceedings of the National Academy of Sciences (PNAS), scientists at MIT's McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center show that an artificial neural network can build its own learning machinery. Starting from random connections and following only simple local rules about how and when connections should change, the network reshaped itself into an effective information-processing system. For more context on this story, see our ongoing AI trends.
Why Biological Plausibility Matters
"Current AI technology is working very well, but it doesn't necessarily use a learning algorithm that is used by the brain," said Qianli Liao, a K. Lisa Yang ICoN Center postdoctoral fellow in the labs of MIT professors Tomaso Poggio and Mark Harnett, and co-first author of the paper alongside MIT research scientist Liu Ziyin and graduate student Yulu Gan.
The distinction matters to neuroscientists who want to use AI models as windows into the brain. Modern machine learning relies on back-propagation, an algorithm most neuroscientists consider unlikely to exist in biological neural circuits. "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."
The team's central argument is that a network's value as a model of the brain should be judged by the extent to which it can organize itself — not just its connections, but the machinery that adjusts those connections.
How the Self-Organizing Network Works
The researchers built networks whose structure begins as multiple layers of randomly interconnected, neuron-like units. Within each layer, every unit is governed by the same simple rules for how connections should be adjusted. "They just interact with each other using that rule, and they develop some type of behavior," Liao explained.
The rules are the kind neuroscientists believe rewire brain circuits during learning — including the classic principle that "neurons that fire together wire together," meaning simultaneous activation strengthens the connection between two neurons.
As the network learned, self-assembling motifs emerged according to these simple rules and then served as building blocks for functional circuits. Those motifs went on to reorganize themselves using an optimization strategy resembling gradient descent — the error-driven adjustment process built into most modern machine learning — but arising from the network's own local dynamics rather than an externally imposed algorithm.
Performance Comparable to Standard Networks
The crucial test was whether this self-organized system could actually learn. It could: the team showed that a self-organizing network governed by these local rules learned to recognize patterns and classify images, performing at a level similar to conventionally designed networks trained through back-propagation.
Senior author Tomaso Poggio framed the result as a potential bridge between two fields that have long been chasing the same mechanisms from opposite directions. If the synaptic motifs the team proposes are what the brain uses to accomplish what back-propagation does for artificial networks, he noted, researchers would have "a strong connection at the level of fundamental mechanisms between artificial and natural networks — between machines and brains."
From Brain-Computer Interfaces to Better AI
The implications run in both directions. For neuroscience, understanding the simple rules by which circuits self-organize could clarify how learning works — and how it breaks down. Liao points to brain-computer interfaces, which aim to restore communication for people who have lost language abilities or let paralyzed patients control prosthetic limbs, as a field where knowing these rules is essential. "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." The same knowledge could inform therapies for conditions in which learning rules are disrupted by aging or disease.
For machine learning, the findings suggest an underexplored design space. Liao cautions that a self-organizing network does not necessarily learn the way the brain does — but he frames self-organization as a critical starting point. "Self-organization is the first high-level framework," he said. "Once we are good with self-organization as a measure, then we can talk about next thing." The group's main objective is understanding the brain, but the knowledge could ultimately help engineers build networks that perform better or behave in more human-like ways.
There is also an open question about granularity. The rules that reshape neural connections as an animal learns may differ across brain regions, and they may change as the brain matures — a possibility the researchers say neuroscientists must map out before simple-rule models can be matched to specific biological circuits. That mapping, in turn, could reveal where and how the learning machinery is disrupted by aging or disease, informing prevention strategies as well as therapies.
The study, from the McGovern Institute for Brain Research and the K. Lisa Yang ICoN Center, was covered by MIT News on October 13, 2026.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
