The Pulse
MIT Network Builds a Learning Algorithm From Random Connections
MIT researchers report that locally changing connections can organize a randomly connected neural network into a system that approximates gradient descent. The computer model performed at a level comparable to backpropagation-trained networ
AI.info Team · Written with AI ·
A neural network that starts with random connections can organize itself into a system that approximates gradient descent, MIT researchers report in a paper published online October 9. The model’s performance matched backpropagation-level results, according to the authors, offering a computational account of how learning machinery might arise from local interactions rather than a designer specifying it in advance. The work does not establish that the brain uses this mechanism. MIT News described the study on October 13.
From random connections to learning circuits
Most artificial neural networks begin with an architecture and a training method chosen by people. The MIT team instead asks whether simple rules for changing connections can help a network develop the machinery it needs to learn. Their model starts with randomly connected, neuron-like units arranged in layers, then lets those units adjust their connections through local interactions.
The paper calls the key building block a Self-Assembling Motif, or SAM. Under heterosynaptic plasticity rules, a motif emerges from initially random connections; networks of those motifs then organize hierarchically into dynamics that approximate a generalized form of stochastic gradient descent. Gradient descent improves a model by adjusting parameters in response to errors, while backpropagation is the method commonly used to calculate those adjustments across a network.
“This suggests that biological learning need not be prescribed but can emerge from local rules—and to a far greater extent than previously thought,” write Qianli Liao et al., authors of “How biological synapses self-assemble gradient learning”.
Comparable performance, not proof about brains
MIT says the network learned to recognize patterns and classify images, performing at a level similar to networks trained with backpropagation. That comparison matters because neuroscientists generally consider standard backpropagation an unlikely description of how biological brains learn: it depends on a globally coordinated calculation, while synapses change through local processes.
The researchers’ proposal offers a possible bridge between those scales. Repeated local changes could assemble motifs that collectively carry out a learning process resembling gradient descent. The result is a model-based demonstration, not evidence that biological circuits have been observed assembling SAMs or using them to learn.
Qianli Liao, a postdoctoral fellow at MIT’s K. Lisa Yang Integrative Computational Neuroscience Center and a co-first author, told MIT News that engineers typically design the rules that govern connection changes in artificial networks. “The brain doesn't have an external designer or architect,” he said. “The brain has to organize itself. And we need to know how simple rules are going to make the brain organize.”
The test is in biological circuits
The distinction sets the study’s next scientific hurdle: whether comparable motifs or learning dynamics can be identified in living neural circuits. The paper presents its framework as a source of testable predictions about cortical organization, not a finding that resolves how the brain learns. MIT’s account also stresses that a self-organizing network need not learn in exactly the same way as a brain.
The researchers say a better understanding of the rules behind neuron-to-neuron interactions could inform neuroscience questions about learning and, eventually, brain-computer interfaces. For now, the concrete result is narrower: a computational network built from local plasticity rules can assemble motifs that approximate gradient-based learning and reach performance comparable to backpropagation systems.