Type-II neural symmetry detection with Lie theory.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Symmetry detection is vital for AI explainability and efficiency.
• Existing methods may not fully capture complex transformations.
• Lie theory offers a robust mathematical framework for symmetry analysis.

Purpose of the Study:

• To develop a neural symmetry detection approach using Lie theory.
• To enable efficient application of symmetry transformations in a latent space.
• To provide insights into data's geometric structure via transformation magnitude estimation.

Main Methods:

• Data projection into a low-dimensional latent space.
• Leveraging the matrix exponential for affine and non-affine transformations.
• Estimating transformation magnitude distributions.

Main Results:

• Demonstrated effectiveness in detecting complex symmetries on augmented MNIST.
• Accurate capture of both affine and non-affine transformations.
• Identification of structural priors aligned with data symmetries.

Conclusions:

• The proposed method enhances AI interpretability and parameter efficiency.
• Potential applications include improved data augmentation and model selection.
• This work advances the understanding of data geometry for AI.