How artificial neural networks build their own learning machinery

Researchers at MIT's McGovern Institute for Brain Research and K. Lisa Yang ICoN Center report that an artificial neural network can construct its own learning machinery, according to MIT News. In a paper published Oct. 9 in PNAS, they argue that a network's capacity to organize itself, without a developer dictating how it learns, should be a key test of whether it can tell us anything about the brain. Qianli Liao, Liu Ziyin and Yulu Gan are co-first authors, working with MIT professors Tomaso Poggio and Mark Harnett.

The model starts as layers of randomly connected neuron-like units, with every unit in a layer following the same rules for changing connections. These include the principle that neurons active at the same time strengthen their link. The team says small motifs assemble themselves under these rules and become building blocks for working circuits, which then reorganize in a way resembling gradient descent. The network learned to recognize patterns and classify images, performing about as well as networks trained with back-propagation, the standard AI method that neuroscientists consider unlikely to operate in the brain.

Liao says understanding such rules matters for studying aging and disease and for brain-computer interfaces, and could eventually improve machine learning too. Poggio says that if the brain really uses these synaptic motifs to do what back-propagation does for AI, it would tie machines and brains together at a fundamental level. Liao cautions that self-organization does not prove a network learns as the brain does, describing it as a first framework for judging biological plausibility.