Critical Percolation as a Synthetic Data Model for Interpretability
Researchers have developed a new synthetic dataset model called critical percolation, designed to better reflect the hierarchical structure of natural data for interpretability studies in neural networks. This model generates sparse, low-dimensional fractal clusters with a power-law size distribution, offering analytical tractability and known critical exponents. The data generation process is efficient, enabling large-scale sampling and analysis, and initial experiments show that the ground-truth latent variables can be decoded from neural network activations. AI
IMPACT Provides a more realistic testbed for interpretability research, potentially leading to better understanding and debugging of neural networks.