MNIST Binary Classification — Neural Network From Scratch vs Keras (Grid Search + PSO)

Binary image classification on MNIST digits 0 vs 1 using:

• A fully-connected neural network built from scratch with NumPy (manual forward/backprop).
• A TensorFlow/Keras baseline with the same architecture.
• Hyperparameter optimization using Grid Search and Particle Swarm Optimization (PSO).

Course: Theoretical Foundation of Machine Learning (Cairo University, 2025 – First Semester).
Instructor: Abdelrahman Sayed.
Report included in this repo.

---

Project Overview

The pipeline follows:

1. Load MNIST and filter classes {0,1} (binary setting).
2. Normalize pixel values to [0, 1].
3. Flatten images to a 1D feature vector.
4. Train:
   • NumPy neural network (from scratch)
   • Keras model (baseline)
5. Optimize hyperparameters:
   • Grid Search
   • PSO
6. Compare results + generate plots.

---

Results (from report)

• From-scratch model

  • Train Accuracy: 0.9993
  • Validation Accuracy: 0.9992
  • Test Accuracy: 0.9991

• Keras model

  • Train Accuracy: 0.9988
  • Validation Accuracy: 0.9976
  • Test Accuracy: 0.9991

• Optimization

  • PSO reached 1.0000 validation accuracy (best discovered configuration)

---

Visualizations

Keras training curves (loss & accuracy)

Validation comparison (from-scratch vs Keras)

If your filenames differ, update the paths above.

---

Repository Structure

. ├─ src/ │ └─ nn_from_scratch_vs_keras.py ├─ assets/ │ ├─ keras_training_results.png │ └─ models_comparison.png ├─ report/ │ └─ ML_Course_ProjectReport.pdf ├─ requirements.txt └─ README.md

---

Setup

1) Create and activate a virtual environment

Windows (PowerShell):

python -m venv .venv
.venv\Scripts\Activate
macOS/Linux:
python -m venv .venv
source .venv/bin/activate

2. Install dependencies pip install -r requirements.txt Run From the project root:

python src/nn_from_scratch_vs_keras.py The script:

Runs grid search over predefined configurations,

Prompts you to enter a custom network configuration,

Trains a from-scratch NN and a Keras NN,

Runs PSO optimization,

Generates plots (saved as PNG files).

Notes This repo intentionally does not include any virtual environment folders (e.g., tf_env/, .venv/).

The task is simple (0 vs 1), so both models converge quickly and achieve near-perfect accuracy.

Authors Omnia Ali Mohamed

Amina Mohy Eldeen

Basma Maher

Kenzy Hamdy

Shahd Abdelalim

License This project is for educational use. If you reuse parts, please provide attribution.