Theoretical investigation on why neural networks fail to learn shape perception
This code implements data generatino, experiments and analyses for the paper Jarvers & Neumann (in preparation). Teaching deep networks to see shape: Lessons from a simplified visual world.
We use docker to manage dependencies and make the experiment runs reproducible.
To build the docker container, you can run docker build -t shapelearning . in
the project root directory.
Each experiment is provided as a Python script in the experiments folder.
For example, to run the first experiment, go to that folder and run
python experiment1_bias.py
Parameters of the experiments can be controlled with command line arguments. For example, to run the first experiment on all combinations of shapes, colors / textures, and image sizes, run:
python experiment1_bias.py --shape rectangles --pattern color --imgsize small
python experiment1_bias.py --shape rectangles --pattern striped --imgsize small
python experiment1_bias.py --shape LvT --pattern color --imgsize small
python experiment1_bias.py --shape LvT --pattern striped --imgsize small
python experiment1_bias.py --shape rectangles --pattern color --imgsize large
python experiment1_bias.py --shape rectangles --pattern striped --imgsize large
python experiment1_bias.py --shape LvT --pattern color --imgsize large
python experiment1_bias.py --shape LvT --pattern striped --imgsize large
The results of each run are stored in subdirs of experiments/figures.