MNIST Digit Recognition: MLP vs. CNN

This project is a practical comparison of two different neural network architectures for the task of classifying handwritten digits using the MNIST dataset. It demonstrates the evolution from a standard dense network (Multilayer Perceptron) to a more advanced Convolutional Neural Network (CNN).

📖 Project Overview

The goal of this project is to demonstrate the practical differences in implementation, data preparation, and performance between standard dense layers and convolutional layers using TensorFlow/Keras.

1. Dense Model (MLP)

The initial approach uses a classic feed-forward structure.

• Preprocessing: 2D images are "flattened" into 1D vectors of 784 pixels ($28 \times 28$).
• Architecture: Two fully connected (Dense) hidden layers with 512 and 256 neurons using ReLU activation.
• Optimizer: RMSprop.
• Output: 10 neurons with Softmax activation representing digits 0-9.

2. Convolutional Model (CNN)

The second part implements a spatial-aware architecture that preserves the 2D nature of images.

• Preprocessing: Data is reshaped to (28, 28, 1) to maintain spatial hierarchy.
• Architecture: * 3 Convolutional layers (Conv2D) with 32 and 64 filters to extract visual features (edges, curves).
  • Max Pooling layers (MaxPooling2D) for spatial reduction and shift invariance.
  • A Flatten operation followed by a Dense classifier head.
• Optimizer: Adam.

📊 Performance Comparison

Model Architecture	Epochs	Batch Size	Test Accuracy
Dense (MLP)	10	128	~98.1%
Convolutional (CNN)	10	64	~99.2%

MLP results:

CNN results:

Key Finding: The CNN model consistently outperforms the MLP by better understanding the spatial relationships between pixels, making it more robust to distortions.

🛠️ Tech Stack

• Python 3
• TensorFlow / Keras - Deep Learning framework.
• NumPy - Mathematical operations on arrays.

🚀 How to Run?

1. Clone the repository:

   git clone https://github.com/Mr-TwisT/CNN-MNIST.git

2. Install dependencies:

   pip install tensorflow numpy

3. Run the script:

   python MNIST-nn.py

🧠 Key Concepts Covered

• Data Normalization: Scaling pixel values to the [0, 1] range.
• Convolutions & Pooling: How "filters" extract features like edges and curves.
• Softmax Activation: Turning raw network outputs into interpretable percentages.
• Flattening: Bridging the gap between 2D feature maps and 1D classifiers.