MaskTune: Mitigating Spurious Correlations by Forcing to Explore

ABSTRACT:

This article examines techniques for mitigating deceptive patterns in deep learning models, drawing inspiration from a novel approach known as MaskTune. MaskTune is a novel solution to tackle the problem of models having an excessive number of parameters in comparison to the available training data. It aids the models in acquiring significant patterns without excessively fitting into insignificant details. MaskTune employs a masking technique to mitigate the model's over-reliance on a limited number of characteristics. By concealing certain characteristics, the model is compelled to uncover new ones. This masking process is implemented in a singular fine-tuning phase, which is a crucial step in adjusting a pre-trained model to a different task or dataset. This methodology reduces both time and computational resources required while enhancing task performance in comparison to training a novel model from the beginning. In addition, a job of selective categorization was developed, using MaskTune's capability to promote resilient learning that is less reliant on deceptive characteristics. This enables the model to identify the absence or concealment of crucial characteristics and prevent generating incorrect forecasts. To assess the efficacy of MaskTune in this job of selective categorization, precise metrics were used to measure the accuracy of predictions made by the model and its capacity to abstain from making choices when there was insufficient information for dependable predictions.

Purpose:

This repository hosts the code developed for the Neural Networks exam project. The objective, based on revising the solution proposed in the seminal paper titled MaskTune: Mitigating Spurious Correlations by Forcing to Explore, was to investigate and implement techniques aimed at minimizing the impact of spurious correlations in deep learning architectures."

The repository contains code for training and evaluating various neural network models on the CIFAR-10 and CelebA datasets. Follow the steps below to use the code.

How to Use the Code

1. Clone the Repository

Start by cloning the repository to your local machine:

git clone https://github.com/Saad-data/masktune_models.git

2. Install Required Python Packages

Navigate to the project directory and install the necessary Python packages listed in requirements.txt:

pip install -r requirements.txt

3. Download and Prepare the CIFAR-10 Dataset

No manual download is needed for the CIFAR-10 dataset. It will be automatically downloaded and saved in the project's data folder.

4. Train the VGG Model on CIFAR-10

To load and train the VGG model on the CIFAR-10 dataset, run the vgg.py script:

python src/vgg.py

You can customize the training process by modifying parameters like batch_size, num_epochs, learning_rate, and momentum within the script.

5. Evaluate the VGG Model

To evaluate the accuracy of the VGG model before and after fine-tuning, run the evaluate_vgg.py script:

python src/evaluate_vgg.py

This will provide the network's accuracy on the entire dataset and on each individual class.

6. Run Selective Classification for the VGG Model

To perform selective classification using the VGG model, execute the selective_classification_vgg.py script:

python src/selective_classification_vgg.py

7. Download and Prepare the CelebA Dataset

1. Download the img_align_celeba folder from this link.
2. Extract the contents, and you should see a folder named archive.
3. Move this folder into the project's data/celeba directory to allow the network to load the dataset correctly.

8. Train the AttentionMaskingResNet50 Model on CelebA

To load and train the AttentionMaskingResNet50 model on the CelebA dataset, run the resnet50.py script:

python src/resnet50.py

You can customize the training process by modifying parameters like batch_size, num_epochs, learning_rate, momentum, and num_classes.

To train the model on a subset of the dataset for computational resource reasons, adjust the subset_size variable.

9. Run Selective Classification for the AttentionMaskingResNet50 Model

To perform selective classification using the AttentionMaskingResNet50 model, execute the selective_classification_resnet.py script:

python src/selective_classification_resnet.py

Repository Structure

• data/: Contains dataset files.
• src/: Contains source code files.
• requirements.txt: Lists the Python packages required for the project.

Notes

• Ensure you have the necessary computational resources to train the models, especially for large datasets like CelebA.
• Modify the script parameters according to your preferences and computational constraints.

License

This project is licensed under the MIT License - see the LICENSE file for details.