(A) study on the computational principles of brain function using artificial deep neural network

Abstract

Our brain has an innate ability to perform various cognitive functions by processing external sensory stimuli. These functions are known to be processed by neurons that selectively respond to specific sensory information, but it is unclear how such properties emerge and what functions they have. Recently, as artificial deep neural networks (DNNs) are showing human-level performance in various perceptual tasks, DNNs are suggested as a new test bed for understanding the computational principles of cognitive functions. In this thesis, by using network models such as artificial DNNs, we study the computational principle of the visual/auditory functions in the brain. We found that basic perceptual abilities observed in the brain – perceiving number of items in a visual scene

or to detect music from other sounds

can also emerge in DNN models, even without training about these functions. Our results suggest that both the process of generalization of sensory data from external worlds and the hard-wired structure of neural networks can provide the basis of innate perceptual functions in the brain.