Which low-dimensional kernel makes the map breathe?
low-D kernel
step 0 · KL = —
Interactive KL-divergence optimization for t-SNE A fixed target similarity matrix P is compared with a low-dimensional similarity matrix Q. Repeated optimization steps move nine embedded points, decrease KL divergence, and align Q with P. Attraction emphasis Current embedding Y y₁ y₂ Current Q(Y) The pairwise probability distributions cell pairs, sorted by target similarity probability target P current Q
Cauchy mode brakes near-neighbor collapse while keeping moderate-distance repulsion alive.
Teaching comparison: P is fixed. Cauchy mode adds a visible near-attraction brake; exact t-SNE uses the Cauchy kernel in Q and its gradient.