MNIST Digit Classification

This repository contains a Jupyter Notebook that implements and compares three ML approaches for the MNIST handwritten-digit dataset: the Perceptron Learning Algorithm (PLA), Softmax (logistic) Regression, and a Linear Regression baseline.

Project Structure

• digit_mnist_classification.ipynb — main notebook with data loading, training, evaluation, and visualization.

Requirements

• Python 3.8+ recommended
• Packages used in the notebook:
  • numpy
  • scikit-learn
  • matplotlib
  • tqdm

Install the packages with pip:

python -m pip install --upgrade pip
python -m pip install numpy scikit-learn matplotlib tqdm

If you plan to run in Google Colab, most packages are preinstalled; you can still run the pip command in a Colab cell to ensure latest versions.

How to run

Local Jupyter (recommended):

# install requirements (see above)
python -m pip install numpy scikit-learn matplotlib tqdm
# start notebook server
jupyter notebook digit_mnist_classification.ipynb

Open the notebook and run cells sequentially. The notebook is annotated and divided into logical sections: data prep, PLA, Softmax, Linear Regression, and evaluation.

Google Colab:

Use the badge or open directly: Open in Colab

Configuration

Key top-level variables in the notebook that control training:

• NUM_OF_ITERATION_PLA — iterations for the Perceptron training loop.
• NUM_OF_ITERATION_SOFTMAX — iterations for Softmax gradient descent.

Change those variables near the top of the notebook to adjust trade-offs between runtime and accuracy. Note: higher iteration counts increase training time significantly.

Results & Evaluation

The notebook computes per-model accuracy, confusion matrices, and a summary table of accuracy vs. training time. See the Summary section inside the notebook for example historical results (different iteration counts tested).

Notes and Tips

• The MNIST dataset is fetched with sklearn.datasets.fetch_openml — this requires internet on first run.
• Training PLA in a one-vs-all setup can be slow for large iteration counts; consider reducing NUM_OF_ITERATION_PLA for quick experiments.

License

This repository currently has no formal license. If you intend to share or publish the code, consider adding a license (for example, MIT). If you want, I can add an LICENSE file with MIT text.

---

If you'd like, I can also add a minimal requirements.txt, an explicit LICENSE file, or insert the summary table from the notebook into this README. Tell me which of those you'd like next.