Integrative activity of neural networks may code virtual spaces with internal representations

•Spatial differentiation of fMRI activity indicates its integrative nature.•Limitations of neuroimaging techniques preclude the direct interpretations.•Using random numbers, we trained neural networks to integrate.•Hierarchical networks learned faster and more efficiently than a single network.•Spatial differentiation of the model activity reproduced the presented stimulation. It was shown recently in neuroimaging that spatial differentiation of brain activity provides novel information about brain function. This confirms the integrative organisation of brain activity, but given present technical limitations of neuroimaging approaches, the exact role of integrative activity remains unclear. We trained a neural network to integrate information using random numbers so as to imitate the “centre-periphery” pattern of brain activity in neuroimaging. Only the hierarchical organisation of the network permitted the learning of fast and reliable integration. We presented images to the trained network and, by spatial differentiation of the network activity, obtained virtual spaces with the presented images. Thus, our study established the necessity of the hierarchical organisation of neural networks for integration and demonstrated that the role of neural integration in the brain may be to create virtual spaces with internal representations of the objects.