Modern artificial intelligence systems can recognize faces, generate human-like dialogue, and write functional code, but they achieve these feats through a design process that bears little resemblance to how the brain learns. While engineers meticulously craft the rules that govern how an artificial neural network adjusts its connections during training, the brain must organize itself without an external architect. A team of researchers at MIT’s McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center has now demonstrated that an artificial neural network can autonomously assemble its own learning infrastructure, using the same kind of simple, local rules thought to operate in biological circuits. In a paper published October 9 in the journal PNAS, the scientists argue that the ability to self-organize is a critical benchmark for determining whether a machine-learning model can genuinely inform neuroscience.
Why Self-Organization Matters for Neuroscience
The fundamental challenge in using AI to understand the brain is that current neural networks rely on learning algorithms designed by humans. Engineers define precisely how connection strengths should change when the network processes data and receives feedback about its errors. This approach works remarkably well for practical applications, but it offers limited insight into biological learning because the brain has no external developer. “The brain doesn’t have an external designer or architect,” says Qianli Liao, a K. Lisa Yang ICoN Center Postdoctoral Fellow and co-first author of the study alongside MIT Research Scientist Liu Ziyin and graduate student Yulu Gan. “The brain has to organize itself. And we need to know how simple rules are going to make the brain organize.”
The path from random connections to functional circuits may be far simpler than anyone imagined.
Neuroscientists study how the rules that reshape neural connections differ across brain regions and how they change with aging or disease. This knowledge is essential for developing therapies for neurological disorders and for technologies such as brain-computer interfaces that restore communication or control of prosthetic limbs. Liao emphasizes that achieving these goals requires understanding the fundamental principles of neuron-neuron interaction, not just designing circuits that work well. An artificial neural network’s value to neuroscience should therefore be judged by the degree to which it can self-organize its own learning machinery, not just its connection strengths.
How a Random Network Learns Without a Programmer
To test whether self-organization is feasible, Liao, Ziyin, and Gan, working with MIT professors and McGovern Institute members Tomaso Poggio and Mark Harnett, constructed a computational network that begins with multiple layers of randomly interconnected neuron-like units. Within each layer, every unit follows an identical rule about when and how to adjust its connections. “They just interact with each other using that rule, and they develop some type of behavior,” Liao explains. The rules are inspired by Hebbian plasticity, often summarized as “neurons that fire together wire together”—when two connected units are activated simultaneously, the strength of their connection increases.
The team demonstrated that this self-organizing network, governed only by these local Hebbian rules, can learn to recognize patterns and classify images. As the network processes data, it spontaneously forms assemblies of strongly interconnected units—what the researchers call “motifs”—that serve as building blocks for functional circuits. Critically, these motifs reorganize themselves in a way that mathematically resembles gradient descent, the optimization strategy at the heart of most modern machine learning. In gradient descent, connections are gradually adjusted in small steps to reduce the error between the network’s predictions and the correct answers. The self-organizing motif network achieves a similar effect without requiring a centralized error signal or an explicit back-propagation algorithm.
The self-assembled network performed at a level comparable to standard artificial neural networks that use back-propagation, the algorithm that many neuroscientists consider biologically implausible because it requires information to flow backwards through the network’s layers. By showing that a network can approximate gradient descent purely through local, biologically plausible rules, the MIT team provides a bridge between machine learning and neurobiology.
What Self-Supervised Learning in Neural Networks Actually Means
When people ask how neural networks can learn without human-labeled data, they are often referring to self-supervised learning, where a network generates its own supervisory signal from the structure of the data itself. The MIT study goes a step further by addressing self-organization of the learning algorithm. In this framework, the network not only learns from data but also builds the mechanism that dictates how learning occurs. The researchers show that with minimal instructions about when connections should strengthen or weaken, a collection of random units can reshape itself into a system that effectively processes information and improves its performance over time. This is distinct from conventional self-supervised learning, which still relies on a human-designed optimization rule. The key finding is that the rule for learning itself can emerge from simple local interactions.
Implications for Bridging Brains and Machines
If the brain uses synaptic motifs analogous to those that emerge in the MIT network, this would establish a deep mechanistic connection between artificial and natural intelligence. “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,” says Tomaso Poggio.
Such a connection could accelerate progress in both fields. For neuroscience, it provides a testable hypothesis about how local plasticity rules might combine to implement global optimization. For artificial intelligence, understanding how the brain organizes its learning infrastructure could inspire more efficient, adaptable, and robust machine-learning systems. Liao notes that self-organization is only the first high-level framework. “Once we are good with self-organization as a measure, then we can talk about the next thing.” The group’s main objective remains understanding the brain, but Liao acknowledges that the same insights could ultimately lead to AI systems that perform better or in more human-like ways.
The Practical Stakes: From Brain-Computer Interfaces to Disease Models
The search for biologically plausible learning rules is not merely an academic exercise. Brain-computer interfaces, which aim to restore lost communication or movement, require a precise understanding of how neurons interact during learning. “None of these can work without knowing the rules behind neuron-neuron interactions,” Liao says. “That’s not achievable by just designing a circuit that works very well.” Similarly, models of aging and neurodegenerative disease need to capture how learning rules change over time. If the self-organizing motif framework proves valid, it could provide a platform for simulating disruptions to these rules and testing potential interventions.
Looking Beyond Human Supervision
The MIT team’s work challenges the assumption that powerful learning algorithms must be designed from scratch. Instead, it suggests that the basic building blocks of intelligence—the rules for adjusting connections based on co-activity—may be sufficient for a network to bootstrap its own learning apparatus. This does not mean the network learns exactly like the brain, but it establishes a necessary condition for biological plausibility. The next steps include scaling the self-organizing approach to more complex tasks and testing whether additional local rules can produce even richer learning capabilities.
As artificial intelligence grows more capable, the question is not merely whether machines can outperform humans on specific tasks, but whether their learning strategies can illuminate the fundamental principles that govern our own minds. By showing that a network can autonomously assemble its learning infrastructure with no more guidance than a few basic rules, the MIT researchers have opened a new avenue for exploring that question. The path from random connections to functional circuits may be far simpler than anyone imagined.
- Why does self-organization matter for neuroscience?Self-organization matters because the brain has no external designer, so understanding how simple local rules enable neural networks to organize themselves is essential for using AI to study biological learning.
- How does a random network learn without a programmer?The network starts with random connections and uses Hebbian plasticity rules, where neurons that fire together wire together, to autonomously develop functional learning circuits without human programming.
