Handwritten Digit Recognition Using Neural Network

This repo builds a 3-layer neural network from scratch to recognize the MNIST Database of handwritten digits, only based on a python library numpy.

Getting Started

The example implements these concept:

• Weight Initialization
• Layer
  • ReLu
  • Softmax
• Training
  • Backpropagation
  • Gradient Descent
• Loss
  • Cross-entropy Loss
  • Regularation Loss

Running the Codes

python main.py

In the main.py, you can modify the learning rate, epoch and batch size.

python test.py

Use test.py to show the test images by opencv and print the predicted result.

Results

• batch size: 1
• learning rate: 0.001
• epoch: 5
• final loss: 0.6

• accuracy: 0.92

Blog Post

https://medium.com/deep-learning-g/handwritten-digit-recognition-using-neural-network-67d7ec76a013