MIT Team Shows Random Networks Can Self-Organize Into Learning Machines, in PNAS
MIT News AI · rss · 2026-10-14
Researchers at MIT's McGovern Institute argue in a PNAS paper that an artificial neural network's value to neuroscience hinges on its ability to self-organize without an external designer, since the brain has no architect.
- Starting from randomly connected units governed by simple local rules (e.g., Hebbian 'fire together, wire together'), the network spontaneously reshapes into an effective information-processing system.
- Self-assembling synaptic motifs emerge during learning and act as building blocks, converging on an optimization strategy akin to gradient descent — while avoiding back-propagation, which neuroscientists consider biologically implausible.
- The self-organizing network matches backprop-designed networks on pattern recognition and image classification.
The authors say the framework could inform research on brain learning rules, aging, disease, and brain-computer interfaces; Tomaso Poggio notes that if these motifs mirror how the brain implements backprop-like computation, it would tightly link artificial and natural networks.
More from Research
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23