Far from being randomly organised, neurons in the primary visual cortex of mice form synaptic connections with each other according to a complex set of wiring rules. Specifically, the way neurons connect depends on the visual stimulus features that they respond to – for example, their selectivity for the orientation or direction of motion of a visual stimulus. Advances in large-scale techniques for measuring activity in the brain are revealing the details of these functional connectivity patterns, but it's less clear exactly why these patterns emerge in the first place.
A new paper published in Current Biology has demonstrated that these connectivity rules may emerge from a simple underlying principle. The team comprising DPhil student Sebastian Klavinskis-Whiting, former MSc student Emil Fristed, former DPhil student Yossi Singer, former Royal Commission for the Exhibition of 1851 Fellow Dr Florencia Iacaruso, now a group leader at the Francis Crick Institute, Professor Andrew King and Associate Professor Nicol Harper showed that the functional specificity of the connections found in the mouse visual cortex arise in artificial neural networks trained to predict the next frame of a movie from its recent past.
This principle – termed 'temporal prediction' – could account for the various ways in which the visual feature preferences of cortical neurons have been shown experimentally to relate to their wiring biases. Crucially, the better the network was at temporal prediction, the closer its own connectivity patterns matched those of the real mouse brain. Moreover, artificial networks trained for different visual tasks – such as compressing their inputs or removing noise – showed a much weaker correspondence with the functional specificity of connections observed in the brain. Overall, these findings suggest that the connectivity patterns of primary visual cortex are optimised to support temporal prediction.
This research expands our understanding of the mouse brain by providing a unifying explanation for the functional (how cells respond) and structural (how cells connect) properties of the primary visual cortex. This work highlights how surface-level complexity in the brain may emerge from simpler underlying principles.
The finding that these simple neural networks naturally organise themselves like the brain suggests that there are inherent commonalities between the brain and artificial neural networks. This work also demonstrates that considering the connectivity patterns of neurons, rather than just their response properties, is a valuable approach for assessing computational models of the brain.
Sebastian Klavinskis-Whiting comments, 'Our research shows that the intricate connections between visual neurons can result from a surprisingly simple principle. I'm excited to see how we can apply these networks to other brain areas to guide future experimental research.'
Read the paper here