Recently, the accuracy has been improved by using "deep learning", which is a basic technique of artificial intelligence, to decipher this brain activity.On the other hand, the operation of deep neural circuits is so complicated that it is difficult to explain "why the decoder gave the answer to a given data".

　Researchers at Okayama University, Rikkyo University, National Institute for Physiological Sciences, Araya Co., Ltd., and Keio University have developed a new method to intuitively explain the decoding of brain activity by the deep neural circuit that is a black box like this. bottom.This was achieved by combining a technique called anti-real virtual explanation and another technique of deep learning called a deep generative model.

　Specifically, based on the brain activity data that the decoder made a mistake in answering, a virtual brain activity data that looks exactly like the real thing is generated by a deep generative model by hostile learning.If the decoder can correctly judge from the virtual brain activity data, the brain region that caused the wrong answer can be determined by comparing the original brain activity data with the virtual brain activity data.It is possible to visually show the difference between the two and give an "explanation" of the malfunction.

　We also succeeded in extracting the characteristics of brain activity used by deep neural circuits for decoding by generating brain activity data with exaggerated characteristics by a brain activity generator and having the decoder read it.These results are expected to be useful as basic techniques for realizing brain image diagnosis using "explainable" deep learning.

Paper information:[Frontiers in Neuroinformatics] Counterfactual Explanation of Brain Activity ClassifiersUsing Image-To-Image Transfer by Generative Adversarial Network