⚡ A Convolutional Neural Network from Scratch

Emma's network for the dermatologists has 105,411 parameters: two convolutions with max-pooling, a dense layer and a softmax output. The Keras notebook builds it in a few lines. This notebook writes every layer by hand with NumPy, forward and backward, checks the gradients numerically and trains it with mini-batch gradient descent on the same synthetic lesions.

The interactive notebook includes:

The convolution by hand, with the example of the book's figure.
Padding, and the convolution as a single matrix product (im2col).
Max-pooling, and where each maximum was.
The forward and backward pass of every layer.
The book's network and its 105,411 parameters.
A numerical check of all the gradients.
Training with mini-batch gradient descent, validation and early stopping.
The evaluation on the test images and the filters the network has learned.

Want to modify the code and experiment on your own?
Open the notebook in Google Colab

Bibliography

LeCun, Y., Bottou, L., Bengio, Y. and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324. https://doi.org/10.1109/5.726791
Goodfellow, I., Bengio, Y. and Courville, A. (2016). Deep Learning. MIT Press. Chapter 9, convolutional networks. https://www.deeplearningbook.org/
He, K., Zhang, X., Ren, S. and Sun, J. (2015). Delving deep into rectifiers: surpassing human-level performance on ImageNet classification. ICCV 2015. The initialization used for ReLU layers. https://arxiv.org/abs/1502.01852