1
$\begingroup$

While the original paper of t-SNE is based on the SNE and SNE uses $\sigma_i^2$ (note the subscript $i$) while calculating the similarity of point $x_i$ WRT other points $j$. i.e., for each data point $x_i$, the corresponding $\sigma_i^2$ is used. But t-SNE uses only $\sigma^2$ (not $\sigma_i^2$) while calculating the similarity of every data point centred at $x_i$. This is even mentioned in the original paper of t-SNE (below Equation 3).

However, I found a lot of resources on t-SNE that claim $\sigma_i^2$ for data point centred at $x_i$ (not $\sigma^2$), which is completely contrary to what is used in the paper. However, I found this lecture mentioning the mathematics of t-SNE correctly (as used in the original paper).

What is actually going on?

$\endgroup$

1 Answer 1

1
$\begingroup$

t-SNE calculates different sigma for each data point, thus the sigma should be indexed. See https://en.wikipedia.org/wiki/T-distributed_stochastic_neighbor_embedding for details.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.