Comparison of projections from multiple data sets using UMAP, UMAP in tensorflow, parametric UMAP, parametric UMAP with an autoencoder loss, parametric t-SNE, t-SNE, SCVIS, IVIS, PHATE, a VAE, an AE, and PCA. (a) Moons. (B) 3D buffalo. (c) MNIST, (d) Cassin's vireo song segments, (e) Mouse retina single-cell transcriptomes. (f) Fashion MNIST, (g) CIFAR10. The Cassin's vireo data set uses a dynamic time warping loss and an LSTM network for the encoder and decoder for the neural networks. The image data sets use a convnet for the encoder and decoder for the neural networks. The bison examples use a t-SNE perplexity of 500 and 150 nearest neighbors in UMAP to capture more global structure.