Affine Geometry of Gaussian ReLU Networks via Conditional Kac-Rice Formulas
What it is
This paper derives exact finite-width formulas for the expected number of activation switches and scalar kinks in ReLU networks using conditional Kac-Rice theory. It demonstrates that training redistributes affine complexity rather than simply contracting it, validated on the Breast Cancer Wisconsin dataset.
The problem it solves
Lack of precise analytical tools to quantify how the piecewise-linear geometry of ReLU networks evolves during training.
What you could use it for
- Analyzing the geometric complexity of neural network decision boundaries
- Predicting the number of affine regions in a trained ReLU network
- Understanding how supervised learning affects network piecewise-linear structure
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.