Neural Network Tutorials

This repository contains a set of Jupyter/Colab notebooks for practicing core neural-network workflows in PyTorch, comparing them with classic machine-learning baselines, and exploring fairness-related datasets. The notebooks move from simple synthetic examples to image classification and tabular fairness experiments.

Notebook Overview

NeuralNetworkBasics.ipynb

Introductory PyTorch experiments on synthetic data from scikit-learn.

Coverage:

• Built a linear regression model with nn.Linear and trained it with mean squared error.
• Built a logistic regression classifier with sigmoid activation and binary cross-entropy loss.
• Implemented a small feed-forward neural network classifier and a deeper neural network.
• Compared neural-network results against DecisionTreeClassifier and RandomForestClassifier.
• Practiced the basic PyTorch loop: tensor conversion, forward pass, loss calculation, backpropagation, optimizer step, and evaluation.

Saved outputs show strong performance on the synthetic tasks, including low linear-regression MSE and classification accuracies around the high-80% to low-90% range depending on the model.

CelebA.ipynb

Computer-vision tutorial using the CelebA face-attributes dataset from torchvision.

Coverage:

• Downloaded and loaded CelebA with torchvision.datasets.CelebA.
• Resized images to 128x128, converted them to tensors, and normalized them.
• Inspected the 40 CelebA attribute labels.
• Built a convolutional neural network with convolution, batch normalization, ReLU, max-pooling, and fully connected layers.
• Trained the model as a multi-label classifier using BCEWithLogitsLoss.
• Evaluated accuracy per attribute and average accuracy across all 40 attributes.

The saved run reports an average attribute accuracy of about 89.5% after a short training run.

FairnessModel.ipynb

Tabular classification experiment based on the Adult Income dataset.

Coverage:

• Loaded the Adult dataset in Colab from an uploaded adult.data file.
• Assigned readable column names for demographic, work, education, and income fields.
• Cleaned missing rows and one-hot encoded categorical variables.
• Converted the processed tabular data into PyTorch tensors and DataLoaders.
• Built a fully connected neural network to predict income class.
• Trained and evaluated the model while tracking training and test losses.
• Saved model and optimizer checkpoints under a results/ directory during training.

This notebook practices neural networks on demographic tabular data and sets up the groundwork for thinking about protected attributes such as sex and race in prediction tasks.

FairJob.ipynb

Fair ranking/click-prediction experiment using Criteo's FairJob dataset from Hugging Face.

Coverage:

• Loaded criteo/FairJob from Hugging Face with pandas.
• Explored columns such as click, protected_attribute, senior, displayrandom, rank, categorical features, and numeric features.
• Split the dataset into train and test sets.
• Converted the tabular data into PyTorch tensors and DataLoaders.
• Built a feed-forward neural network to predict whether a job impression was clicked.
• Trained and evaluated the click model with negative log-likelihood loss.

This notebook focuses on applying PyTorch to a fairness-oriented ranking dataset and experimenting with the mechanics of learning from protected-attribute and ranking features.

Tools And Libraries

The notebooks primarily use:

• Python
• PyTorch and Torchvision
• scikit-learn
• pandas
• NumPy
• tqdm
• Jupyter Notebook / Google Colab

How To Run

Open the notebooks in Jupyter or Google Colab and run the cells from top to bottom. Some notebooks require downloading or uploading data:

• CelebA.ipynb downloads CelebA through Torchvision, which is a large dataset.
• FairnessModel.ipynb expects the Adult Income adult.data file to be uploaded in Colab.
• FairJob.ipynb reads the FairJob dataset from Hugging Face.

These notebooks are tutorial experiments rather than a packaged Python project, so there is no shared application entry point.