Building and Evaluating Neural Networks for MNIST

Overview

This project focuses on building, training, and evaluating several neural network architectures on the MNIST dataset of handwritten digits using PyTorch. The goal is to compare the performance of different architectures, from a simple single hidden layer to a convolutional neural network (CNN), and analyze their accuracy and learning behavior.

Project Structure

deep_learning/
├── data/                    # MNIST dataset (downloaded automatically)
├── code_for_project.ipynb   # Main Jupyter notebook (training + evaluation)
├── report_of_project.pdf    # Full report with methods, results, and appendices
└── README.md                # Project documentation

Report

For detailed explanations, training results, and full code listings, see the report.

How to run

Requirements

Make sure you have Python ≥ 3.9. You can install dependencies using:

pip install torch torchvision matplotlib pandas

Run the notebook

Open the Jupyter notebook and execute all cells:

jupyter notebook code_for_project.ipynb

The code will:

1. Download the MNIST dataset automatically.

2. Train and evaluate three neural network models.

3. Print and plot the accuracy after each epoch.

Technical details

Libraries used

• torch, torchvision
• matplotlib
• pandas

Datasets

• MNIST: 60,000 training and 10,000 test grayscale images (28×28 pixels).

Architecture implemented

Model	Description	Accuracy
Single hidden-layer NN	1 hidden layer (400 neurons)	≈ 92.5 %
Two hidden-layer NN	Layers of 500 and 300 neurons	≈ 97.99 %
Convolutional NN (CNN)	2 conv layers + dense layers	≈ 99.1 %

Each model is trained using SGD with momentum = 0.99 and learning rate = 0.001. Loss function: CrossEntropyLoss.

Results

• Increasing the network depth improves performance up to ~99 % accuracy.

• CNN achieves the best trade-off between speed, generalization, and accuracy.

• Training curves and complete implementation details are available in the report.

Key learnings

• How network depth impacts model accuracy.

• Practical use of PyTorch for training and evaluating neural networks.

• Visualization and normalization of image data.

• Hyperparameter tuning (learning rate, momentum, weight decay).

Author

Clara Petrescu-Stompor (clara.petrescu.stompor@gmail.com)