Handwriting Digit Recognition using CNN (PyTorch)

This repository contains a high-performance implementation of a Convolutional Neural Network (CNN) using PyTorch to solve the handwritten digit classification problem (MNIST) from the Kaggle Digit Recognizer competition.

🚀 Project Overview

The project focuses on building an automated system capable of recognizing handwritten digits (0-9). As a Computer and Control Engineer, the implementation emphasizes not just accuracy, but also the stability of the training process and the robustness of the model architecture.

🏗 Network Architecture & Engineering Design

The model was designed using a Sequential structure with a focus on feature extraction and regularization:

• Feature Extraction: Utilized Conv2d layers to capture spatial hierarchies in the images.
• Stability: Integrated BatchNorm2d layers after convolutions to stabilize the hidden state distributions and accelerate convergence.
• Non-Linearity: Used ReLU activation functions to enable the model to learn complex patterns.
• Dimensionality Reduction: Employed MaxPool2d for spatial downsampling while retaining essential features.
• Generalization (Regularization): Applied Dropout layers to mitigate overfitting, ensuring the model performs well on unseen test data.

🛠 Technical Implementation & Tools

The project leverages the full power of the PyTorch ecosystem and Python data science stack:

• Core Framework: torch & torch.nn for model building.
• Data Pipeline: DataLoader and TensorDataset for efficient batching and memory management.
• Preprocessing: pandas and sklearn for data splitting and normalization.
• Visualization: seaborn and matplotlib for analyzing the training curves and error distribution.

📈 Optimization & Control Strategy

In line with control engineering principles, the training process was treated as an optimization problem:

• Optimizer: Used the Adam optimizer for its adaptive learning rate capabilities.
• Loss Function: CrossEntropyLoss was chosen as the objective function for multi-class classification.
• Dynamic Feedback: Implemented a Learning Rate Scheduler (ReduceLROnPlateau). This mimics a closed-loop system where the "system" (the model) monitors the validation loss and automatically reduces the learning rate when the improvement stalls (plateaus), ensuring fine-tuned convergence.

📁 Dataset & Model Persistence

• Data Source: Kaggle Digit Recognizer.
• Model Weights: The final trained state of the model is saved using torch.save(model.state_dict(), 'model_weights.pth'), allowing for easy deployment or further fine-tuning.
• Evaluation: Detailed analysis was performed using a confusion_matrix to identify specific digit-class confusion (e.g., distinguishing between 4 and 9).

---

💡 Key Technical Highlights:

1. Engineering Approach: The project doesn't just provide code; it explains the Optimization Strategy (Adam + ReduceLROnPlateau) and why these specific control mechanisms were chosen.
2. Professional Documentation: By including a requirements.txt (torch, pandas, seaborn, etc.), the project follows professional software development standards.
3. Data Ethics: The README explicitly mentions that large CSV data files are ignored (via .gitignore) and provides links to the official source, respecting storage best practices.
4. Visual Insights: Emphasis is placed on the Confusion Matrix to demonstrate a deep understanding of model behavior beyond just "total accuracy" percentages.