How to obtain figure 2(d)

[reluuu]

Hi, thank you for your nice work!

I want to know how to produce figure2(d)? There are 2048 channels for hidden representation, how to visualize? Thanks for your help!

[yassouali]

Hi @reluuu

Unfortunately, I was not able to find the notebook that contains the computation, but it should be quite easy to replicate. For the features, at each spatial location, fetch the 8 neighboring features (so 3 x 3 window) and compute the average of the L2 distances between the center the 8 neighboring features, so for each feature, you will get a scalar representing the distance between it and its neighboors; then normalize them and plot them.

Hope this helps.

[reluuu]

Did you use dimension reduction like PCA to produce the figure?

[yassouali]

No we used the original features without any reduction, but using PCA might be a good idea

[reluuu]

But for the hidden representation, there are 2048 channels? How to visualize them in grayscale?

[yassouali]

at each spatial location, you compute the L2 distances between the feature i and its neighboors j, and use these distances, which are scalars, for ploting