Artificial neural networks can develop specialized internal modules as they learn multiple tasks, according to research published in Nature Machine Intelligence on September 28.
The study offers a computational explanation for one feature of brain organization: groups of connected units that specialize while remaining part of a larger network.
Researchers trained recurrent neural networks on cognitive tasks and compared how different learning schedules affected their organization. Networks learning several tasks became more modular than those trained on a single task, particularly when the demands placed on them strained their available capacity.
Introducing tasks incrementally produced the strongest modular organization while maintaining superior performance. These networks also developed structural features that more closely resembled biological brain networks than models shaped only by spatial constraints.
The findings suggest that the work a network must perform can help shape its architecture. That adds a functional explanation alongside theories emphasizing physical factors, such as the cost of connecting neurons.
The experiments concern artificial networks, however, and do not establish that they think like humans. Their value lies in showing how task demands can produce organized specialization through learning.