Synthetic datasets

Setup

Ensure that the root of this repository is on your PYTHONPATH environmental variable. For the digit classifciation example, we also need Morpho-MNIST downloaded and on the PYTHONPATH.

See .env for an example PYTHONPATH.

Spiral classification

Classic two-dimensional nonlinear classification benchmark. Data points are sampled along two (or more) spiral functions and the neural network must learn to classify points by their spiral. Random effects are simulated by dividing the data into clusters and random varying the spiral radius in each cluster.

See spiral_classification.ipynb for a comparison of conventional and mixed effects models. Spiral generation parameters can be varied to control the degree of random effects and, optionally, confounding effects.