Year
2025
projecT Duration
IN PROGRESS
Developed a 2 layer neural network from scratch for classifying handwritten digits using the MNIST dataset, achieving 95% accuracy through effective architecture design and parameter tuning.
Applied Xavier initialization to weights and biases for improved convergence and implemented forward and backward propagation using ReLU activations in hidden layers and softmax for output.
Evaluated performance using cross-entropy loss and accuracy metrics. Currently integrating the Adam optimizer and tuning hyperparameters to accelerate training and prevent overfitting.
Clean and modular code structure written in Python, focusing on transparency and educational value.
Architecture: 1 hidden layer (e.g., 128 neurons), input layer (784), and output layer (10 classes).
Manual tuning of hyperparameters such as learning rate, batch size, and epochs to avoid overfitting and ensure stability.
Implemented the full training pipeline, including forward and backward propagation.
Used ReLU activation in hidden layers and softmax at the output layer.
Employed Xavier initialization to improve weight convergence.
Trained using cross-entropy loss. Currently integrating the Adam optimizer for better generalization and faster convergence.
Python for model implementation
NumPy for vectorized math operations
Matplotlib for visualizing performance metrics
Jupyter Notebook for experiments and visualization
