Table 1. Summary of Reweighted Stochastic Embedding Methods Used in (Section 3.2): mrse and stke^a.

	mrse	stke
reweighting factor r(xk, xl)	[eq 29]	[eq 31]
high-dim. prob. M(xk, xl)	Gaussian mixture [eq 30] for perplexities ∈ {256, 128, 64, 32}	Gaussian [eq 32] with ε = 0.12
low-dim. prob. Q(zk, zl)	t-distribution [eq 26]	Gaussian with ε = 0.12
landmark sampling	weight-tempered random sampling for τ = 3 and 5000 landmarks	minimal pairwise distance for rc = 1.2 and 97000 landmarks
activation functions	ReLU (3 layers)	hyperbolic tangent (3 layers)
optimizer	Adam (μ = 0.001, β1 = 0.9, β2 = 0.999)	Adam (μ = 0.001, β1 = 0.9, β2 = 0.999)
batch size	1000	256

^aLabels: reweighting factor r(x_k, x_l), high-dimensional Markov transition matrix M(x_k, x_l), and low-dimensional Markov transition matrix Q(z_k, z_l).