Scientists at MIT's McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience Center argued in a paper published Oct. 9 in PNAS that a key metric for an artificial neural network's biological plausibility is whether it can organize itself without a developer's guidance. They reported that a network of randomly connected units, given simple rules about how and when connections should change, could reshape itself into an effective information-processing network and learn from its training data.

The network began with layers of randomly interconnected neuron-like units governed by uniform rules within each layer. The team reported that self-assembling motifs emerged and served as building blocks for functional circuits, and that the network learned to recognize patterns and classify images. It performed at a similar level to networks using back-propagation, an algorithm the researchers said neuroscientists consider unlikely to be used by the brain.

Co-first author Qianli Liao said a self-organizing network does not necessarily learn the way the brain does, but called self-organization a first high-level framework. The authors framed the work as a step toward understanding brain learning, with possible long-term relevance to brain-computer interfaces and machine learning; they did not report clinical results.