MNIST Digit Recognition (KNN vs CNN)

Project Overview

This project implements handwritten digit recognition using the MNIST dataset and compares two different approaches:

• K-Nearest Neighbors (KNN) as a classical machine learning baseline
• Convolutional Neural Network (CNN) as a deep learning model

The goal is to classify digits (0–9) and analyze performance differences between traditional ML and deep learning methods.

Dataset

• MNIST handwritten digit dataset
• 70,000 grayscale images
• Image size: 28 × 28 pixels
• 10 classes (digits 0–9)

Method 1: K-Nearest Neighbors (KNN)

Description

A simple instance-based learning algorithm used as a baseline classifier.

Steps

• Load MNIST dataset
• Normalize pixel values
• Split into train/test sets
• Train KNN classifier (k=3)
• Evaluate performance

KNN Result

• Accuracy: 97.13%

Method 2: Convolutional Neural Network (CNN)

Description

A deep learning model designed for image classification using convolutional layers.

Architecture

• Conv2D (32 filters)
• MaxPooling2D
• Conv2D (64 filters)
• MaxPooling2D
• Flatten
• Dense (64 neurons)
• Output Layer (Softmax, 10 classes)

Steps

• Load MNIST dataset
• Normalize images
• One-hot encode labels
• Train CNN model
• Evaluate performance

CNN Result

• Test Accuracy: 99.11%

Comparison

Model	Accuracy
KNN	97.13%
CNN	99.11%

Key Insight

CNN outperforms KNN due to:

• Better feature extraction (convolutions)
• Spatial awareness in images
• Ability to learn hierarchical patterns

Technologies Used

• Python
• NumPy
• Scikit-learn
• TensorFlow / Keras

How to Run

KNN Model

python knn_classifier.py

CNN Model

python cnn_classifier.py

Conclusion

This project demonstrates that deep learning (CNN) significantly outperforms traditional machine learning (KNN) for image classification tasks such as handwritten digit recognition.

👩‍💻 Author

Fateme Khosravi Computer Science Graduate | Interested in Data Science, Algorithms, and Systems Analysis