The tsne plot of Generated features are not coinciding with original features using GAN

vision
1

Hi,

I am using GAN for feature generation, but the tsne plot of generated one is not overlapping or nearby the original features.

x1 -> x2

x1 (original feature)
x2 (generated feature)
Can anyone suggest the solution?
Thanks

2

Disclaimer: I’m really not sure what you’re trying to do :). I assume you have two approaches for dimensionality reduction (t-SNE and GAN) and try to compare the results.

Well, t-SNE is a probabilistic approach where the distances between data points in the original space and the lower-dimensional space get represented by (conditional) probabilities, the optimization goal is to minimize the difference between this two resulting probability distributions.

That’s probably net really enlightening but the important part is (a) different runs of t-SNE will give you different results – depending in the random initialization, and (b) t-SNE does not guarantee to preserve things like cluster sizes, orientations, shapes or things you might preserve with linear approaches such as PCA or LDA.

In short, it’s not obvious to me why the generation features of your 2 methods should overlap.

1 Like
3

Thanks @vdw for giving the more insights. Actually I am trying to see the orientations of generated samples (using GAN) with the original features. And after several experiments I found that GAN is not properly working (not able to generate features properly). So to check the orientations, I used the tsne plots. But yeah you are right. I just want to use GAN for feature generation from one domain to other.
Thanks