Deep Generative Neural Embeddings for High Dimensional Data Visualization

Anonymous

Keywords: Deep Generative Models, Embedding

TL;DR: Deep Generative Neural Embeddings for High Dimensional Data Visualization

Abstract: We propose an embedding-based visualization method along with a data generation model. In particular, corresponding locations of data points in the visualization are optimized as embeddings along with a generative network such that the network could reconstruct the original data. The generalization aspect of the neural network enforces similar data points to be close in the embedding space. Since our method includes the generative part, it allows visualizations that are not possible with neighborhood embedding methods such as TSNE. Compared to parametric methods such as VAE our method is non-parametric and relaxes the need to optimize the encoder part, allowing us to obtain better optimizations.