🔢 MNIST Digit Classification using ANN

Recognizing handwritten digits with a fully connected neural network built in TensorFlow and Keras

---

📌 Overview

This project builds an Artificial Neural Network (ANN) that classifies handwritten digits, 0 through 9, using the classic MNIST dataset. The full workflow lives inside a single Colab notebook and covers everything from raw pixel data to a trained model with 97.43% test accuracy.

It is a clean example of how a relatively simple feed forward network, with the right preprocessing and a touch of dropout regularization, can perform extremely well on a well structured image classification problem.

---

✨ Highlights

🎯 What	📋 Detail
Final Test Accuracy	97.43%
Dataset Size	42,000 labeled digit images
Input Features	784 pixels per image (28 x 28 flattened)
Architecture	4 layer fully connected ANN with dropout
Training Epochs	10
Framework	TensorFlow / Keras Sequential API

---

🧠 How It Works

flowchart LR
    A[📂 Load mnist.csv] --> B[🔍 Explore data]
    B --> C[⚙️ Normalize pixels 0 to 1]
    C --> D[🏷️ One hot encode labels]
    D --> E[✂️ Train test split]
    E --> F[🏗️ Build ANN model]
    F --> G[🚀 Train for 10 epochs]
    G --> H[📊 Evaluate accuracy]
    H --> I[🖼️ Visualize predictions]

Loading

---

🗂️ Dataset

The dataset is the classic MNIST handwritten digits dataset, provided here as a flat CSV file (mnist.csv) rather than the usual image format.

• 🧾 42,000 rows, each representing one digit image
• 🎨 785 columns total, one label column plus 784 pixel0 to pixel783 columns
• 🔢 Each pixel value ranges from 0 to 255, representing grayscale intensity
• 🧩 Every row, once reshaped, forms a 28 x 28 grayscale image

Before training, the pixel values are scaled down to a 0 to 1 range, and the digit labels are converted into one hot encoded vectors for the 10 output classes.

---

🏗️ Model Architecture

A simple yet effective Sequential ANN built with Keras:

model = Sequential([
    Dense(256, activation='relu', input_shape=(784,)),
    Dropout(0.3),
    Dense(128, activation='relu'),
    Dropout(0.3),
    Dense(64, activation='relu'),
    Dense(10, activation='softmax')
])

Layer	Type	Units	Activation	Notes
1	Dense	256	ReLU	Input layer, takes 784 flattened pixels
2	Dropout	—	—	30% dropout for regularization
3	Dense	128	ReLU	Hidden layer
4	Dropout	—	—	30% dropout for regularization
5	Dense	64	ReLU	Hidden layer
6	Dense	10	Softmax	Output layer, one node per digit

⚙️ Compilation settings

• 🧮 Optimizer: Adam
• 📉 Loss function: categorical_crossentropy
• 📈 Metric tracked: accuracy

---

🚀 Training

The model was trained for 10 epochs with a batch size of 64, using an 80/20 train test split.

Epoch	Train Accuracy	Val Accuracy	Train Loss	Val Loss
1	75.48%	94.60%	0.7642	0.1741
5	96.70%	97.18%	0.1029	0.0931
10	98.18%	97.43%	0.0591	0.0942

📊 Both accuracy and loss curves for training and validation are plotted at the end of training, making it easy to visually confirm the model is learning well without overfitting heavily.

---

✅ Results

🏆 Final Test Accuracy: 97.43%

After training, the model is evaluated on the untouched 20% test split. The notebook also picks the first 5 test images and displays each one alongside its true label and the model's predicted label, giving a quick visual sanity check on real predictions rather than just a single accuracy number.

---

🛠️ Tech Stack

Category	Tools Used
🐍 Language	Python 3
🧠 Deep Learning	TensorFlow, Keras
📊 Data Handling	Pandas, NumPy
📈 Visualization	Matplotlib
🤖 Preprocessing	Scikit learn (train_test_split)
☁️ Environment	Google Colab

---

📁 Project Structure

MNIST-by-using-ANN-model/
│
├── mnist (1).ipynb     # Main notebook: data loading, training, evaluation
└── README.md           # You are here

---

▶️ Getting Started

1. Clone the repository

   git clone https://github.com/sarthakNaikare/MNIST-by-using-ANN-model.git

2. Open the notebook Upload mnist (1).ipynb to Google Colab or run it locally in Jupyter.

3. Get the dataset The notebook expects a file named mnist.csv at /content/mnist.csv. Place your copy of the MNIST CSV dataset there before running.

4. Run all cells Execute the notebook top to bottom. Training takes well under a minute per epoch on Colab's standard runtime.

---

🔮 Future Improvements

• 🧬 Swap the ANN for a CNN to push accuracy even higher
• 🎛️ Add hyperparameter tuning (learning rate, layer sizes, dropout rate)
• 🛑 Introduce early stopping and learning rate scheduling
• 🧪 Build a small confusion matrix view to spot which digits get confused most often
• 🌐 Wrap the trained model in a simple web demo for live digit recognition

---

👤 Author

Swapnil Yadav

⭐ If you found this project useful or interesting, consider giving it a star!