An artificial neural network can be judged by how well it performs a task. But if researchers want to use such systems to understand the brain, performance may not be enough. A team of scientists at MIT argues that a more telling question is whether a network can organize itself, including the machinery that changes its connections as it learns.
Why self-organization matters
Many AI systems learn by adjusting the strengths of connections between artificial neurons. That approach can be powerful, but a network whose structure and learning process are designed from the outside may tell neuroscientists little about how a brain learns. Brains do not have an external designer or architect to arrange their circuits.
The researchers say models should therefore be assessed by how much they can organize on their own. That includes both the connections among units and the mechanisms that alter those connections. The distinction matters because a model can solve a task without showing that the way it learned resembles biological learning.
The rules that change connections may differ across brain regions or shift as the brain matures. Understanding those rules, and how aging or disease might disrupt them, could help researchers develop therapies and prevention strategies. The same knowledge is relevant to brain-computer interfaces intended to restore communication or enable paralyzed patients to control neuroprosthetic limbs.
A network that develops its own circuits
Working with colleagues, Liao and his team created a computational network made of multiple layers of randomly connected, neuron-like units. Units within a layer followed the same basic instructions for adjusting their connections. Rather than specifying a finished circuit, the researchers gave the network rules and allowed its components to interact.
One rule reflects the idea that “neurons that fire together wire together.” When connected units activate at the same time, their connection grows stronger. Rules of this kind are thought to help reshape brain circuits as people and other animals learn from experience.
As the network learned, small connection patterns assembled and reorganized. These motifs became building blocks for circuits that could process information. The team reports that the model learned to recognize patterns and classify images, showing that simple local rules could produce useful behavior without a developer laying out the final network.
How the approach compares with modern AI
The researchers found that the emerging motifs reorganized through a process resembling gradient descent. In this approach, connections are adjusted in response to errors so that performance can improve. Most modern machine-learning systems use gradient descent through some form of back-propagation, a method neuroscientists consider unlikely to be used by the brain.
The team compared its self-organizing network with networks designed to learn through that approach. The self-organizing model performed at a similar level. That result suggests a network may be able to learn effectively while developing the structures that support its learning, rather than relying on a fully specified design.
Similar performance does not show that the brain uses the same process. The researchers present self-organization as a starting point for judging biological plausibility, not as proof that their model reproduces how natural brains learn. The open question is whether the motifs proposed by the team, or related mechanisms, are actually present in the brain.
What the findings could mean
If researchers find that biological circuits use mechanisms comparable to those in the model, it could connect ideas about learning in artificial and natural networks. That connection could inform neuroscience and AI alike. For now, the work offers a framework for asking whether an artificial network can develop its own learning machinery.
The immediate goal is to better understand the brain. Liao also points to a possible longer-term benefit for machine learning: insights into self-organization could help engineers build systems that perform better or learn in more human-like ways. The research, published Oct. 9 in PNAS, makes that possibility a question for further study rather than a settled outcome.