Parametric UMAP Embeddings for Representation and Semisupervised Learning.

Area of Science:

• Machine Learning
• Data Science
• Computational Topology

Background:

• Uniform Manifold Approximation and Projection (UMAP) is a powerful nonparametric method for dimensionality reduction.
• UMAP leverages Riemannian geometry and algebraic topology to create low-dimensional embeddings of complex data structures.
• The standard UMAP algorithm involves graph construction and stochastic gradient descent optimization.

Purpose of the Study:

• To extend the UMAP algorithm by introducing a parametric optimization approach using neural networks.
• To enable fast, online embeddings for new data points through a learned parametric mapping.
• To explore the application of parametric UMAP as a regularization technique and for enhancing semisupervised learning.

Main Methods:

• Developed a parametric extension of the UMAP algorithm by optimizing neural network weights.
• Implemented stochastic gradient descent for optimizing the low-dimensional embedding based on neural network parameters.
• Integrated parametric UMAP into autoencoder architectures and semisupervised learning frameworks.

Main Results:

• Parametric UMAP demonstrated performance comparable to the original nonparametric UMAP.
• The learned parametric mapping allows for rapid embedding of new data points.
• Utilizing UMAP as a regularization improved classifier accuracy in semisupervised learning tasks by effectively capturing unlabeled data structure.

Conclusions:

• The parametric UMAP framework offers a computationally efficient and effective alternative to nonparametric methods.
• Parametric UMAP provides a flexible tool for various machine learning applications, including representation learning and semi-supervised classification.
• This research highlights the potential of integrating deep learning with topological data analysis for advanced data embedding and analysis.