k* Distribution: Evaluating the Latent Space of Deep Neural Networks Using Local Neighborhood Analysis.

Area of Science:

• Machine Learning
• Data Visualization
• Artificial Intelligence

Background:

• Traditional dimensionality reduction techniques like t-SNE and UMAP distort local neighborhood structures in neural network latent spaces.
• This distortion hinders the clear visualization and analysis of sample distributions within specific classes.

Purpose of the Study:

• Introduce the k* distribution and its visualization technique to preserve local neighborhood structures.
• Enable detailed analysis of how neural networks process different classes within their latent spaces.
• Provide a more profound understanding of contemporary visualization methods for neural network latent spaces.

Main Methods:

• Developed the k* distribution and a corresponding visualization technique.
• Utilized local neighborhood analysis to ensure the preservation of sample distribution structures for individual classes.
• Compared k* distributions to analyze class processing within neural networks.

Main Results:

• Identified three distinct sample distributions in latent spaces: fractured, overlapped, and clustered.
• Demonstrated that sample distributions significantly vary across different classes within a neural network.
• Showcased the applicability of the k* distribution analysis across diverse network architectures, layers, input transformations, and data splits (training/testing).

Conclusions:

• The k* distribution offers a superior method for visualizing neural network latent spaces by preserving local structure.
• This technique facilitates a deeper understanding of intra-class and inter-class relationships within learned representations.
• The k* distribution is a valuable tool for analyzing and understanding the internal workings of various neural network models.