Machine learning (neural networks, CNNs) has mimicked the human brain. By using it, it became possible to reproduce the intracerebral reaction (MRI) when viewing an image from there! And now you can do the reverse! In other words, the opposite is true, the feature amount of a certain image can be predicted from the reaction result of MRI! That is.
Until now, it was said to be imitation, but I felt that it was a paper that proved that it was not a metaphor but was imitated with really great accuracy!
https://academic.oup.com/cercor/article/28/12/4136/4560155
The convolutional neural network was said to imitate the image processing mechanism of the brain, but it has become clear. In other words, recent research has shown that → fMRI can be created from CNN features, and vice versa! It's amazing that CNN is learning even though it doesn't consider time series! I wondered if end to end is something like this. A paper that reinforces the interactivity between CNN and fMRI. Also, CNN did a good job of predicting not only the tummy side (which I've already predicted perfectly), but also the back side fMRI! Although the degree is low. By directly decoding the fMRI signal, we were able to estimate the characteristic representations of the visual space and the semantic space, and perform visual reconstruction and semantic classification. My opinion: This is just a story that has reached the same level as fMRI, and FMRI should not be completely understood. It seems that it can only be explained in another way that it has become as clear as it is, that is, there is no contradiction between method A and method B. That's it, the convolutional neural network is amazing! I can't say that. There is a possibility, though.
Results on the relationship between CNN and fMRI. The CNN and visual cortex share similar representations of low-level visual features (eg, retinal topi) and high-level semantic features (eg, face), as well as multiple intermediate levels of high abstraction. It also shares a hierarchical representation of visual information (Fig. 2). Results and validity for reproducing nerves. From the result of CNN, I tried to find out which brain region is activated by a linear regression model. Classify by labeling.
Recommended Posts