This project demonstrates various machine learning techniques for classifying handwritten digits from the MNIST dataset. It covers data preprocessing, model training, evaluation, and advanced classification strategies.
- Project Overview
- Setup and Installation
- Data Overview
- Data Visualization
- Model Training
- Model Evaluation
- Multi-Class Classification
- Multi-Label Classification
- Contributing
- License
The project focuses on image classification using the MNIST dataset, which contains 70,000 grayscale images of handwritten digits. We explore binary classification for detecting a specific digit, multiclass classification for all digits, and multi-label classification for scenarios where multiple labels may apply.
To get started, clone the repository and install the required Python libraries. The project uses libraries like NumPy, pandas, Matplotlib, Scikit-learn, and Seaborn for data manipulation, visualization, and modeling. Follow the instructions in the setup script to install dependencies.
The MNIST dataset consists of 60,000 training images and 10,000 test images, each of size 28x28 pixels. Each image is labeled with a digit from 0 to 9. The dataset is well-balanced, though some digits are slightly more frequent than others.
We visualized individual and multiple digits from the dataset to understand the data better. Examples include displaying a single digit as a heatmap and plotting multiple digits in a grid to observe their variations.
Initially, we focused on a binary classification task to detect the digit '5' using the SGDClassifier. We transformed the labels into binary format and trained the model. A dummy classifier was used as a baseline for comparison.
We extended the classification problem to handle all ten digits using the SVC (Support Vector Classifier) and the RandomForestClassifier. Techniques like one-vs-one and one-vs-all strategies were used to manage multiple classes.
We evaluated the models using accuracy, precision, recall, and F1-score. Precision-recall curves and ROC curves were plotted to visualize the trade-offs and overall performance of different classifiers.
Cross-validation was employed to assess the robustness of the models. We used methods like cross_val_score and cross_val_predict to obtain reliable performance metrics and validate the models' effectiveness.
For the MNIST dataset, we implemented both one-vs-one and one-vs-all strategies. We used SVC and RandomForestClassifier to handle multi-class classification, analyzing the performance of each method.
We explored multi-label classification by creating labels for digits greater than or equal to 7 and odd digits. The KNeighborsClassifier was used for this task, and the performance was evaluated using metrics like the F1-score.
Contributions to this project are welcome! Please submit issues, pull requests, or suggestions for improvements. Ensure that your contributions align with the project's goals and coding standards.