Binary Image Classification with Neural Network (NumPy)

This project was completed as part of the "Basics of Deep Learning" course at Colman College, 2024.

We implemented a feedforward neural network from scratch using NumPy to classify grayscale images of hand-sign digits (only digits 1 and 3) from a custom dataset. This assignment included designing the model architecture, implementing forward and backward propagation, and training the model using binary cross-entropy loss.

👨‍💻 Authors

• Tal Levy
• Omer Virshovsky

📊 Dataset

• Hand sign digit images (grayscale 28x28 pixels).
• Subset of 1000 images containing only digits 1 and 3.
• Data was flattened and normalized to values between 0 and 1.
• 80% training, 20% testing split.

🧠 Model Architecture

• Input layer: 784 neurons (flattened 28x28 images)
• 3 hidden layers: 128 neurons each, sigmoid activation
• Output layer: 1 neuron (binary classification)
• Loss: Binary Cross-Entropy (BCE)
• Optimizer: Gradient Descent

📈 Results

• Accuracy: 98% on test set

• Confusion Matrix:

  Actual \ Predicted	1	3
  1	102	2
  3	1	95

• Loss curve:
  (add if you have image)

📁 Files

• basics_2025.ipynb – Jupyter Notebook with full implementation
• report.pdf – Project documentation and discussion
• README.md – You’re here

🧠 What I Learned

• Manual implementation of a neural network using NumPy
• Fundamentals of backpropagation and gradient descent
• Evaluation using metrics such as confusion matrix and accuracy
• Working with real image data and binary classification