Course Project Introduction to Information Technology > Faculty of Information Technology - VNU-HCM University of Science

This project implements and compares the performance of Deep Learning models (MLP & CNN) for handwritten digit recognition using the MNIST dataset. The project is built with a modular architecture, designed for extensibility and academic research purposes

---

📄 Documentation & Paper

The project is accompanied by a detailed scientific report, providing an in-depth analysis of the mathematical foundations and empirical evaluation.

👉 Read the full report (PDF)

---

✨ Key Features

• Modular Architecture: Clear separation between Data Loading, Model, Loss function, and Training loop.
• Multi-Architecture Support:
  • MLP (Multi-layer Perceptron): Fully Connected Neural Network, Basic Feed-Forward Network (Baseline).
  • CNN (Convolutional Neural Network): Convolutional network optimized for spatial feature extraction.
• Reproducibility: Hyperparameter management via yaml configuration files.
• Logging & Visualization: Real-time loss/accuracy tracking and prediction visualization.

📅 Project Timeline and Team Members (HCMUS-ConChoCaoBangBoPC)

Below is the implementation progress and task distribution of the team throughout the development process:

---

📂 Project Structure

digits_classification/
├── docs/                           # Project documentation
│   └── final_article.pdf           # Scientific report file
├── configs/                        # Configuration files
│   └── config.yaml                 # Main config (Epochs, LR, Model type...)
├── assets/                         # Image files
│   └── GanttChart.png              # Project Timeline
│   └── Figure_1.png                # A graph that compares loss over time between the two models during training
│   └── Figure_2.png                # A graph that compares accuracy over time between the two models after every epoch
│   └── Figure_3.png                # An evaluation table for the MLP Model
│   └── Figure_4.png                # An evaluation table for the CNN Model
│   └── cnn_interface.png           # User interface for the CNN Model
│   └── mlp_interface.png           # User interface for the CNN Model
│   └── demo.mp4                    # Interactive Demo
│   └── cnn_confusion_matrix.pdf    # A confusion matrix for CNN Model
│   └── mlp_confusion_matrix.pdf    # A confusion matrix for MLP Model
│   └── mnist_samples_high_res.pdf  # Samples of the MNIST dataset
├── src/                            # Source code
│   ├── dataloader.py               # Data processing module (DataLoader, Transforms)
│   ├── model.py                    # Model architecture definitions (CNN, MLP)
│   ├── trainer.py                  # Training script
│   └── utils.py                    # Utilities (Visualization, reading from config.yaml, etc)
├── model                           # Directory for saving trained model weights
├── main.py                         # Perform all the training, validating, and result visualizing steps defined in 'src/.'      
├── app.py                          # Play a demo of both models
├── requirements.txt                # Project dependencies
└── README.md                       # README file

---

🚀 Installation & Usage

1. Environment Setup

Requires Python 3.8+.

# Clone repository
git clone [https://github.com/username/digits-classification.git](https://github.com/username/digits-classification.git)
cd digits-classification

# Create virtual environment (Recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# venv\Scripts\activate   # Windows

# Install dependencies
pip install -r requirements.txt

2. Training

You can modify parameters in configs/config.yaml before running.

python trainer.py --config configs/config.yaml

The model with the highest accuracy will be automatically saved to saved_models/.

3. Testing:

Evaluate the model on the Test set:

python test.py --model_path saved_models/best_model.pth

4. Interactive Demo (Gradio)

We provide a real-time web interface using Gradio to test both models interactively. This allows for a direct visual comparison of how each architecture handles spatial information.

Launch the Web App

python app.py

Demo Preview

A demonstration of the real-time digit recognition process.

2025-12-27.00-49-02.online-video-cutter.com.mp4

Figure 1: Real-time inference demo showing model predictions.

Side-by-Side Comparison: MLP vs. CNN

As discussed in our analysis, the MLP model often struggles with subtle geometric variations due to its flattened input. In contrast, the CNN leverages convolutional filters to maintain spatial correlations.

Multi-Layer Perceptron (MLP)	Convolutional Neural Network (CNN)
MLP result showing higher ambiguity in probabilities	CNN result showing robust spatial feature extraction

---

👥 Researchers:

HCMUS - CONCHOCAOBANGBOPC - 25CTT3

Members	StudentID
Nguyen Minh Nhat	25120215
Vu Thanh Phong	25120219
Do Le Nhat Quang	25120223
Nguyen Phu Quang	25120224
Nguyen Vu Nhat Quang	25120225
Pham Dang Quang	25120226

Lab Instructor: Mr. Le Duc Khoan.

---

📝 License

This project is distributed under the MIT license.