t-SNE

[Serendipity-LC]

Hi！When we use t-SNE to draw, for the output features of the model, do we use the predicted label of the corresponding model or the true label of the sample？

[PetarV-]

Hello,

Normally we'd build the t-SNE based on the model's hidden representations as inputs, and colour-code it based on the true label. This shows (with caveats! check out https://distill.pub/2016/misread-tsne/ for more details) roughly how well the model is "separating" the classes.

Hope that helps!

Thanks,
Petar

[Serendipity-LC]

Hello,

Normally we'd build the t-SNE based on the model's hidden representations as inputs, and colour-code it based on the true label. This shows (with caveats! check out https://distill.pub/2016/misread-tsne/ for more details) roughly how well the model is "separating" the classes.

Hope that helps!

Thanks,
Petar

This means that predicting labels when drawing pictures is useless for us, right?
Thanks.