Accelerating TSNE with GPUs: From hours to seconds

TSNE (T-Distributed Stochastic Neighbor Embedding) is a popular unsupervised dimensionality reduction algorithm that finds uses as varied as neurology, image similarity, and visualizing neural networks.

Unfortunately, its biggest drawback has been the long processing times in most available implementations. RAPIDS now provides fast GPU-accelerated TSNE, building on the GPU-based Barnes-Hut approach developed at CannyLab.

TSNE in RAPIDS’ cuML machine learning library can run up to 2,000x fasterthan the corresponding CPU implementation (Scikit-Learn) and uses up to 30% less GPU memory than current GPU versions and can be 2x faster. (CannyLab’s BH TSNE).