GALDA-public

This repo contains Matlab scripts and dependencies for the manuscript "Generative Adversarial Linear Discriminant Analysis".

If you would like to run the main GALDA, use GALDA_main_v2.m (which can also be found in cv folder). No need to run the install.m

run demo_adversarialAttack.m for recreating spectroscopic adversarial attack; run demo_GALDA for the GALDA iterative process. (both in demo folder)

GALDA is best run if p >> n, where p is parameters/dimension of the data, and n is the number of experiments (i.e. spectra). It is best if you run when you have 3 or more classes to classify.

---- The main GALDA code will provide 7 figures ----

LDA Loading Vector or Image (w vector (from Fisher Discriminant) used to project the data)
LDA Plot (visual representation of the projection in 2D)
LDA Confusion Matrix
GALDA Resolution J Plot (J vs Iteration Plot - helps visualize overfitting and accuracy of GALDA. It is also the 'resolution of GALDA', so the main idea is to minimize the gap between J_train and J_test (reduce overfitting), while maximizing J_test (accuracy).
GALDA Loading Vector or Image (We assume the first plot corresponds to generated data, while the rest is from the genuine. It lays out by smallest to biggest eigenvalues. You can change to images if twoD = 1)
GALDA Plot (visual representation of the projection in 3D)
GALDA Confusion Matrix

---- Parameters to change ----

Input file name
Wavenumber length (dimension or 'p' of the data)
Number of iteration loops (NumLoops)
Number of generated spectra created each iteration (NumGenSpec - cannot exceed 'p')
K-Fold
twoD

** (You can add colors and markers if needed. Currently, it can go up till 10 classes.)

** Instead of LDAfunc, you can use LDAfuncReg, which requires a regularizer. The standard LDAfunc uses pinv to allow the inverse matrix to work.

** In folder utils, file StatScore, you can change Z. This is the decision boundary. Currently it is set to Euclidean distance, but you can use z-score, Mahalanobis distance, and others if needed to change.

License

Simpson-Laboratory-for-Nonlinear-Optics/GALDA-dev

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Contributors 4

Uh oh!

Languages