ExSpliNet: An interpretable and expressive spline-based neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Interpretable and expressive neural network models are crucial for understanding complex data.
• Existing models may lack interpretability or efficient encoding mechanisms.

Purpose of the Study:

• To introduce ExSpliNet, a novel neural network architecture.
• To demonstrate its interpretability, expressiveness, and universal approximation capabilities.
• To showcase efficient encoding strategies leveraging B-spline properties.

Main Methods:

• The ExSpliNet model integrates concepts from Kolmogorov neural networks, probabilistic tree ensembles, and multivariate B-spline representations.
• A probabilistic interpretation of the model is provided.
• Universal approximation properties are mathematically demonstrated.
• Efficient encoding methods exploiting B-spline characteristics are discussed.

Main Results:

• ExSpliNet exhibits universal approximation properties, indicating its potential for complex function approximation.
• The model's architecture allows for efficient encoding, reducing computational complexity.
• Effectiveness was validated on synthetic approximation tasks and benchmark machine learning datasets.

Conclusions:

• ExSpliNet offers a powerful, interpretable, and expressive alternative for various machine learning applications.
• The model's foundation in B-splines facilitates efficient implementation and understanding.
• Further research can explore its application in diverse scientific and engineering domains.