MIT

MIT researchers show randomly connected neural networks can self-assemble gradient-style learning from local plasticity rules

MIT News published this on Oct 13, 2026 (source: news.mit.edu).

In a peer-reviewed PNAS paper published October 9, researchers at MIT's McGovern Institute and K. Lisa Yang ICoN Center show that layers of randomly interconnected neuron-like units, each following the same simple rules for changing connections, can organize into motifs that carry out a gradient-descent-like optimization without backpropagation. The self-organizing network learned pattern recognition and image classification at a level similar to the backpropagation-trained networks it was compared against. The authors, including Qianli Liao, Liu Ziyin, Yulu Gan, Tomaso Poggio and Mark Harnett, argue self-organization is a key test of biological plausibility for brain models.