Reproduce “SplineCam: Exact Visualization and Characterization of Deep Network Geometry and Decision Boundaries” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ahmed Imtiaz Humayun, Randall Balestriero, Guha Balakrishnan, Richard Baraniuk
SplineCam introduces a provably exact and sampling-free method for visualizing and characterizing the full input-space geometry and decision boundaries of deep networks employing continuous piecewise linear activation functions. This technique precisely computes linear regions and decision boundaries, offering quantitative metrics like Average Region Volume to analyze network behavior, and demonstrates its scalability on large models.
Key insights
- The exact computation of input-space partitions and decision boundaries provides a rigorous foundation for understanding how DNs with CPWL activations transform data.
- Quantitative metrics derived from the partition geometry (e.g., Average Region Volume, Number of Regions) can characterize functional complexity and provide objective comparisons between architectures and training strategies.
- The approach reveals hierarchical learning patterns in Implicit Neural Representations (INRs) and explains the effect of positional encoding on increasing effective linear regions through periodic weight sharing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.