Table 1.
Architecture of the model.
Fig 1.
Diagram of the attention module.
As illustrated, an element-wise summation of each output from the shared network, a sigmoid function activation, an element-wise multiplication between attention and initial features, and an element-wise summation with initial features are performed after the shared network.
Table 2.
Architecture of the generator.
Fig 2.
(A) Validation Accuracy curve (upper) and Validation Loss curve (lower) on the validation set of the visual experiments. (B) Validation Accuracy curve (upper) and Validation Loss curve (lower) on the validation set of the imagination experiments. The chance accuracy is 1/40 = 0.025 (shown by the gray line).
Table 3.
Averaged accuracy of classifying the imagination test set.
Fig 3.
Sorted frequency band of each model.
Sorting was performed based on the attention map averaged over the test set. “Top 1” means the filter with the highest attention value. (A) Frequency band of each spatial filter. (B) Center frequency of each spatial filter.
Fig 4.
(A) Weighted sum of absolute values of spatial filters. Each result was normalized so that the maximum value is 1. (B) FFT results of weighted summed sinc filters.
Fig 5.
The map of the difference between the weighted sums of the spatial filters in the “Mix” model and the “Visual” model which were derived from the visual test set (Left), and between the “Mix” model and the “Imagine” model which were derived from the imagination test set (Right).
Fig 6.
Weighted summed filters in each frequency block.
Each map was normalized so that the maximum value is 1. “Low” block mainly contains alpha band or below, “Mid” block mainly contains beta band, and “High” block mainly contains gamma band.
Table 4.
Numbers of picked filters in each frequency block.
Fig 7.
Results generated by “Visual” GAN.
(A) Sample of images generated from the visual train dataset. (B) Sample of images generated from the visual test set. (C) The classification accuracy map of images generated from the visual train dataset. (D) The classification accuracy map of images generated from the visual test set.
Fig 8.
Results generated by “Mix” GAN.
(A) Sample of images generated from the visual train dataset. (B) Sample of images generated from the visual test set. (C) The classification accuracy map of images generated from the visual train dataset. (D) The classification accuracy map of images generated from the visual test set.
Fig 9.
Results generated from imagination dataset.
(A) Sample of images generated by the “Visual” GAN. (B) Sample of images generated by the “Mix” GAN. (C) The classification accuracy map of images generated by the “Visual” GAN. (D) The classification accuracy map of images generated by the “Mix” GAN.