MIT model shows how neural networks can self-assemble learning circuits
Published in Artificial Intelligence, Brain/Neurology, Journals, Numerical Simulations.
Published October 12, 2026. A PNAS paper dated October 9 presents a computational mechanism through which randomly connected neural units can organize both their circuitry and the machinery that enables learning.
MIT researchers call the approach Self-Assembling Learning. Their proposed building block, the Self-Assembling Motif, emerges through local rules that change synaptic connections. Repeating these motifs across layers produces network dynamics that approximate a generalized form of stochastic gradient descent.
The team reports learning performance comparable to backpropagation in its tested models. MIT’s accompanying report describes pattern recognition and image classification as demonstrations. The work addresses how useful learning circuits could emerge, rather than assuming that those circuits are already in place.
Scope: This is a theoretical and computational result. It does not demonstrate that biological brains use these motifs, and comparable performance in tested tasks does not establish equivalent ability across all AI applications. The model offers hypotheses that neuroscience experiments can test.
Sources: Original PNAS paper, October 9, 2026; MIT McGovern Institute report.