Example of the new staircase structure emerging in algorithms that repeats data.
$h^\star_{sign}=\mathrm{sign}(x_1x_2x_3)$ cannot be learned in$O(d)$ steps, while$h^\star_{stair}=h^\star_{sign}+\mathrm{He}_4(x_1)$ it can because of the staircaise mechanism.
Tested with Python 3.11
git submodule update --init --recursive # install boostmath
pip install -r requirements.txt
pip install -e giant-learning --no-binary :all:
The file structure of this repository is as follows:
giant-learning/contains the Python package used to run the experiments.hyperparameters/contains some example configuration files needed to run the experiments.running.pyis the main script to run the experiments.plotting.pyis the main script to plot the results.example.ipynbis a Jupyter notebook that shows how to run the experiments and plot the results.