This project implements a neural network from scratch to classify MNIST handwritten digits with a high accuracy of 98.02%. It is entirely built using only NumPy without the use of any machine learning libraries like TensorFlow or PyTorch. The network consists of input, hidden, and output layers with backpropagation to adjust weights over 100 epochs.
- Data Preprocessing: Flattened images and normalized pixel values to ensure efficient training.
- Neural Network Layers: Implemented a 3-layer neural network with one hidden layer of 128 units.
- Activation Function: Used the sigmoid activation function for both hidden and output layers.
- Backpropagation: Calculated error gradients and updated weights to minimize the loss function.
- Gradient Descent: Used a fixed learning rate of 0.1 to optimize the model over time.
- Training Accuracy: 98.02%
- Loss: Converges to 0.0007 after 40 epochs