JayRGopal / Explainable-Deep-Learning-Regularizer

A more interpretable way to regularize convolutional neural networks (CNNs) during training

Reg-Explain

Evaluating the impact of regularization on explainability

A custom-built simple CNN has been coded via PyTorch to aid in this endeavor.

Additionally, Resnet50 is being used. It is being trained via code from the PyTorch-Image-Models (TIMM) repository.

The dataset being used is CIFAR10.

Notes

Folders labeled as "OLD" have an old method of formatting that is no longer favored. They remain on the GitHub for reference.

The custom-built transformer has not undergone a thorough explainability analysis yet.

Resnet50 310 Epoch Results

Resnet50-Control: 96.48%

Resnet50-Dropout: 96.34%

Resnet50-L2: 96.42%

Vgg19 310 Epoch Results

Vgg19-Control: 91.45%

Vgg19-Dropout: 91.83%

Vgg19-L2: 91.37%

SimpleCNN 8 Epoch Results

SimpleCNN-Control: 64.00%

SimpleCNN-L2 1e-5: 66.63%

SimpleCNN-L2 1e-8: 65.83%

SimpleCNN-Dropout: 68.32%

SimpleCNN-Dropout-L2 1e-5: 68.06%

SimpleCNN-Dropout-L2 1e-8: 68.76%