MIT study models how learning machinery can assemble itself

A different question about learning

An MIT research team has reported a computational model in which learning machinery forms from simple rules governing local connections. MIT described the work on October 13, following the paper's online publication in PNAS on October 9.

The distinction is between making an already-designed system learn and explaining how a system capable of learning could take shape. The paper, by Qianli Liao, Liu Ziyin, Yulu Gan, Brian Cheung, Mark T. Harnett and Tomaso Poggio, addresses the second question.

Small interactions, larger structure

In the authors' abstract, the proposed mechanism is a Self-Assembling Motif, or SAM: a small connection pattern that emerges from initially random connectivity. Networks of these motifs then organize at a second, larger scale. The authors report that their dynamics approximate a generalized form of stochastic gradient descent.

In plain language, gradient descent makes adjustments aimed at reducing errors. Here, the claim is that an error-reducing learning process can arise through lower-level interactions, rather than every part of its machinery being prescribed separately.

This is not a system without a designer. The model still begins with rules and starting conditions chosen by researchers. Self-assembly describes what emerges within that setup.

Performance is a bounded claim

MIT says the network learned to identify patterns and classify images. The team compared it with networks using backpropagation, a common way to adjust connections during training, and reported similar performance. The abstract also reports a match to backpropagation-level performance.

Neither source supports a claim that this mechanism outperforms frontier language models, makes training generally cheaper or explains every form of learning. A comparison needs its task and conditions attached; removing them turns a research result into a broader claim it has not established.

A model is not a brain

MIT explicitly cautions that a self-organizing network need not learn in the same way as a biological brain. The work offers a candidate mechanism to investigate, not direct confirmation that human neural circuits use it.

This report draws on the authors' abstract and publication metadata, together with MIT's institutional account. The publisher's full text was not accessible in this review, so no detailed benchmark audit or independent reproduction is claimed.

The useful takeaway is narrower than a headline about AI teaching itself: researchers have proposed a way for learning structures to emerge under local rules. Whether that mechanism describes biology remains an open question.