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.
- Train models:
drake %mnist-models
- Wait until all jobs finish and run evaluation:
drake %mnist-results
- Continue at "Figures and tables" section
For row "Innate" table 1:
- Checkout branch
reproduce-innately-local
- Run
train-mnist-innately-local.job
- Execute notebook
reporting/reproduce-innately-local.ipynb
- Train models:
drake %cifar10-models
- Wait until all jobs finish and run evaluation:
drake %cifar10-results
- Continue at "Figures and tables" section
- ReLog activation function figure:
reporting/activation-functions.ipynb
- Pattern fitting figure (two-dot problem):
reporting/two-dots.ipynb
- Regularization figure:
reporting/fitting.ipynb
- Negative examples figure:
reporting/overlay.ipynb
- Activation on noise:
reporting/distributed-vs-local.ipynb
- MNIST and CIFAR-10 result tables:
reporting/results-table.ipynb
- Curvature statistics in Section 4.1:
reporting/curvature.ipynb