G-NeuroDAVIS: A generative model for data visualization through a generalized embedding.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Visualization

Background:

• Visualizing high-dimensional data and generating realistic samples remain significant challenges.
• Existing methods often fail to produce generalized embeddings that capture data structure and generate new data.

Purpose of the Study:

• Introduce G-NeuroDAVIS, a novel generative model for high-dimensional data visualization and sample generation.
• Develop a model capable of creating high-quality, generalized embeddings and realistic high-dimensional samples.

Main Methods:

• G-NeuroDAVIS utilizes advanced generative techniques for effective data representation.
• The model supports both supervised and unsupervised training settings.
• Conditional sample generation is a key feature, assessed qualitatively and quantitatively.

Main Results:

• G-NeuroDAVIS demonstrates superior embedding quality and performance in downstream tasks compared to Variational Autoencoder (VAE).
• The model outperforms VAE, Deep Convolutional Generative Adversarial Network (DCGAN), Denoising Diffusion Probabilistic Models (DDPM), and Autoencoder (AE)-guided Real-valued Non-Volume Preserving (RealNVP) in sample generation.
• Interpolation experiments show smooth, meaningful transitions, indicating preservation of underlying data structure.

Conclusions:

• G-NeuroDAVIS is an effective tool for high-dimensional data visualization and representation learning.
• The model offers significant improvements in generating realistic and diverse samples.
• Its robust performance makes it suitable for diverse applications requiring high-quality data generation.