Deep-learning engine implementation

C++ convolutional neural-network engine implemented from scratch

• Application to MNIST handwritten digit data

  • Data : train-images.idx3-ubyte, train-labels.idx1-ubyte

    • Number of data : 60000 (training: 50000, validation: 10000)

    • Data image resolution : 28x28

  • Training parameters:

    • Layers  and number of neurons : 3, 784-30-10

    • Cost function : Cross-entropy

    • Mini-batch : 10

    • Learning rate : 0.5

    • epochs : 30

  • Result :

    • Recognition ratio : 95.7%

    • Digits to image : success, failure