NeuroMapper: Using UMAP to visualize the training dynamic of neural network

Abstract

Deep neural networks are often considered as black box models. One of the reasons is that they have a large number of parameters. Understanding the model through these parameters is often a hard task. Recent advances in nonlinear dimensionality reduction techniques offer a better way to visualize high-dimensional nonlinear data[1]. These techniques have the potential of helping people understand deep neural networks. However, many of these techniques pose problems, such as indeterministic results and slower computation. These problems prevent them from being used in visualizing the training dynamic of deep neural networks[2]. Neuro-Mapper attempts to tackle this through the use of scalable data visualization technologies and Aligned UMAP, a special kind of UMAP that is able to generate embeddings that are aligned with each other[3]. Users can explore the training behavior of a model using NeuroMapper and adjust the hyperparameter as they see fit. NeuroMapper works across multiple platforms and can visualize more than 40000 data points.