- color_regions.ipynb: Some small experiments with the API used for the data generation process
- color_regions.py: API for data generation process
- finite_differences-colors.ipynb: Experiments with the RGB version of the model/data
- finite_differences-hard.ipynb: Experiments with the "hard" version of the grey scale task + experimetns with the AutoEncoder + graphs of network response to changing random pixels
- finite_differences_easy_greyscale.ipynb: Early experiments, with the "easy" version of the grey scale task (2 classes only)
- finite_differences_medium_greyscale.ipynb: Experiments with "medium" version of task + experiments with PCA directions method + comparison with grad and grad*input methods + experiments on sparseness of map with strides vs. no strides
- mechanistic_interpretation.ipynb: An attempt to explain the algorithm the network learns, under the mechanistic interpretability philosophy, done on a tiny version of the model
Read the paper here