Reproducing results

Notice that the following steps only work on the computing environment they were created for (which is the Dutch cluster Cartesius). Please read the scripts and adapt them to your computing environment.

MNIST experiments

1. Train models: drake %mnist-models
2. Wait until all jobs finish and run evaluation: drake %mnist-results
3. Continue at "Figures and tables" section

For row "Innate" table 1:

1. Checkout branch reproduce-innately-local
2. Run train-mnist-innately-local.job
3. Execute notebook reporting/reproduce-innately-local.ipynb

CIFAR-10 experiments

1. Train models: drake %cifar10-models
2. Wait until all jobs finish and run evaluation: drake %cifar10-results
3. Continue at "Figures and tables" section

Numbers, figures and tables in the paper

1. ReLog activation function figure: reporting/activation-functions.ipynb
2. Pattern fitting figure (two-dot problem): reporting/two-dots.ipynb
3. Regularization figure: reporting/fitting.ipynb
4. Negative examples figure: reporting/overlay.ipynb
5. Activation on noise: reporting/distributed-vs-local.ipynb
6. MNIST and CIFAR-10 result tables: reporting/results-table.ipynb
7. Curvature statistics in Section 4.1: reporting/curvature.ipynb