A deep neural network is an artificial neural network consisting of multiple layers, and has made great progress in the field of artificial intelligence in recent years.

　In the future, when artificial intelligence is used in areas that can be life-threatening, such as autonomous driving and medical diagnosis, it is important to elucidate how animals, including humans, process visual information. ..Therefore, the research team has developed a method to copy the activity of nerve cells in the visual cortex to a visual image to a deep neural network and analyze it in detail on a computer as follows.

　First, the same image as the image shown to the animal is input to the deep neural network, and the activity of the nerve cell with respect to the image is used as a teacher signal to learn the deep neural network.Next, the input image is repeatedly updated so as to maximize the output of this neural network.The resulting image is presumed to be the image that maximizes the activity of the nerve cells copied to the neural network, and is useful for identifying what kind of image input the cells respond to best (reaction selectivity).

　When this method was applied to the analysis of neurons in the primary visual cortex of mice, images corresponding to the reaction selectivity of the primary visual cortex were generated.If this method is applied to higher-order regions, it is expected that unknown reaction selectivity will be identified.

　This method can also be applied to the analysis of other sensory areas, and is useful for elucidating the processing mechanism of sensory information in the brain.Furthermore, by using a deep neural network that captures the activity of brain nerve cells, it is expected that it will lead to the development of artificial intelligence that is closer to the behavior of humans and animals.

Paper information:[Scientific Reports] Characterisation of nonlinear receptive fields of visual neurons by convolutional neural network