Finite Gaussian Neurons for Adversarial Defense

Ph.D. work by Felix Grezes

Cite as

Felix Grezes, Finite Gaussian Neurons for Adversarial Defense, 2021, https://github.com/grezesf/FGN---Research.

@misc{grezes2021,
 author={Felix Grezes},
 year={2021},
 title={Finite Gaussian Neurons for Adversarial Defense},
 howpublished={\url{https://github.com/grezesf/FGN---Research}},
}

Dependencies

PyTorch
Torchvisionn
Numpy
Scipy
Matplotlib

Directory Organization

----\
    |
    |---\Finite_Gaussian_Network_lib
        # functional library to run FGNs
        |
        |---\fgn_helper_lib
            # useful functions not stricly related to FGNs
        |
        |---\tests
            # tests for the library functions
    |
    |---\Notebooks
        # notebooks to plot results, visualize data, etc...
    |
    |---\Experiments
        # contains scripts to run experiments and the results
    |
    |---\dev
        # development work
    |
    |---\old
        # old work

Finite_Gaussian_Network_lib

A collection of functions related to Finite Gaussian Networks.

• Matlab style: one function per file. Open a file to see it's definition, parameters, etc...
• the fgn_helper_lib directory contains randoms useful functions, but that don't directly relate to FGNs.
• the tests directory contains tests for functions in the library

Notebooks

A collection of Jupyter Notebooks used for data visualization, results plotting, experiments analysis, etc... Loosely follows the narrative of the thesis.

Experiments

A collection of tiny scripts that run experiments, and folders containing the results. The scripts should be tiny, only creating the folders, setting the parameters and calling the library function. Convention: scripts should create timestamped folders for the results each run. mnist_fgn_train.py should create `/res-mnist_fgn_train-time:stamp'

Dev

collection of notebooks used to develop the FGN library functions, the scripts.