PNAS Study Shows Artificial Neural Networks Can Self-Assemble Learning Machinery

A study published in PNAS (9 October 2026) demonstrates that randomly interconnected neuron-like units can self-assemble learning motifs using shared plasticity rules. These self-organized networks achieved performance in pattern recognition and image classification comparable to networks trained via backpropagation.

Mechanism and Findings

The research distinguishes between directly designing a learning circuit and specifying local rules that allow the learning machinery to organize itself. The reported comparison highlights that this self-assembly approach yields results similar to traditional backpropagation methods in the tested tasks.

Limitations and Context

The authors explicitly caution that this self-organization does not prove the brain uses this mechanism. The primary motivation is biological plausibility, not evidence that this method replaces backpropagation in large-scale models. Further evaluation is needed regarding the stability of these motifs under different initial connectivity and their extension beyond the demonstrated tasks.