Designing rotationally invariant neural networks from PDEs and variational methods

Area of Science:

• Computer Vision
• Machine Learning
• Mathematical Modeling

Background:

• Partial differential equation (PDE) models and variational formulations often possess inherent rotational invariance, crucial for applications like image analysis.
• Convolutional Neural Networks (CNNs) typically lack this property, necessitating complex workarounds.

Purpose of the Study:

• To investigate the mechanisms of rotational invariance in diffusion and variational models.
• To transfer these principles to design more robust and inherently rotationally invariant neural networks.

Main Methods:

• Proposed novel activation functions that couple network channels by integrating information from multiple oriented filters.
• Ensured rotation invariance at the fundamental building block level of neural networks.
• Maintained the capability for directional filtering.

Main Results:

• Developed neural architectures with inherent rotational invariance.
• Achieved rotation invariance comparable to existing methods requiring extensive orientation sampling, but with significantly fewer and smaller filters.
• Demonstrated a method for translating diffusion and variational models into mathematically sound network architectures.

Conclusions:

• The proposed activation functions provide a mathematically grounded approach to building rotationally invariant CNNs.
• Offers novel concepts for model-based CNN design, enhancing their applicability in fields requiring rotational consistency.