MIT Scientists Build Self-Organizing AI Network That Matches Backprop on Image Tasks
Updated
Updated · MIT News · Oct 13
MIT Scientists Build Self-Organizing AI Network That Matches Backprop on Image Tasks
2 articles · Updated · MIT News · Oct 13
Summary
A PNAS paper published Oct. 9 showed MIT researchers could start with randomly connected neuron-like units and let simple local rules assemble the network’s own learning machinery.
Those rules mimic biological ideas such as “neurons that fire together wire together,” producing self-assembling motifs that reorganize into functional circuits as the system learns patterns and classifies images.
The team said the network reached performance similar to back-propagation-based systems, even though backprop is widely viewed by neuroscientists as an unlikely model of how the brain actually learns.
MIT researchers argue self-organization should be a key test of an AI model’s biological plausibility, because the brain has no external designer shaping its learning algorithm.
That framework could help neuroscience uncover learning rules disrupted by aging or disease and eventually improve brain-computer interfaces and more human-like machine learning.