There are three experiments involved:

1. Modern risk curve for DL: risk_curve.py and risk_curve_evaluate_model.py
2. Measure intrinsic dimensions: intrinsic_dimensions.py and intrinsic_dimensions_measurement.py
3. Measure re-initialization and re-randomization robustness of layers: layer_equality.py

See blog post "Are Deep Neural Networks Dramatically Overfitted?" for more details.