Controlling primate neural activity using artwork from artificial neural networks

By showing macaques images generated by an artificial neural network, researchers were able to control the activity of specific neurons within the visual systems of these animals’ brains, according to a new study. The results demonstrate how the artificial neural networks of today could one day be used by neuroscientists to design novel experiments to study the brain under targeted response conditions, conditions not previously possible to achieve in the brains of living organisms. Artificial neural networks (ANN) – which underlie many computer vision applications – are currently the most accurate computational models of primate vision. This is particularly true for the ventral visual stream, the split-second neural processes that occur between what is observed by the eye and perceived by the brain. While previous research has suggested that visually evoked neural responses predicted by ANNs are quite similar to those observed in living brains, the accuracy of these models’ “understanding” of the ventral stream isn’t well-known; it’s generally believed that they still lack the reliability and performance of primate brains, in vision tasks. To put current computational models of vision to the test, Pouya Bashivan and colleagues used an ANN to control the activity of neurons within the brains of macaques. Using one of the leading ANN ventral stream models, Bashivan et al. synthesized a series of images resembling abstract black and grey patterns, which were predicted by the model to activate specific populations of neurons within a primate brain, when viewed. Viewing the images was also predicted to selectively activate some neuron populations, while leaving others unchanged. The ANN-synthesized images were shown to rhesus macaques equipped with implanted microelectrode arrays used to monitor the monkey’s visually evoked neural responses. The ANN-based approach was able to both selectively predict and independently control neuron populations, the authors say. This represents an ability to control neurons far more accurately than scientists ever have before, they note. In the future, increasingly accurate ANN models could produce even more accurate control of neural activity, they say.

###