Visualizing Data using t-SNE

Introduced t-SNE, resolving the crowding problem of SNE by using a heavy-tailed Student-t distribution in the low-dimensional embedding map.

The classic technique for inspecting high-dimensional neural activation layers, image representations, and word embedding clusters.

Assumptions

• Local topological neighborhoods matter more than long-range metric Euclidean distances for exploratory visualization

Limitations

• Non-parametric: does not learn an out-of-sample projection function for new incoming test points; O(n^2) scaling without Barnes-Hut