Deep learning methods: a practical approach

by Francesco Morandin

Table of contents (tentative)

1. The seventies. Neural networks before computer

Perceptron unit, the biologically inspired artificial neuron

Layer of many units, when categorical output is needed

Multilayer neural networks, a universal approximator with automatic feature extraction

Activation function, the need for some trainable nonlinearity

2. The nineties. Supervised training success

Training set, learning from examples through some loss function

Backpropagation, the CS answer to calculus' chain rule for gradient

Minibatches and stochastic gradient descent

3. The tens. Subtle science and exact art of DL

Autoencoders and unsupervised weight initialization

Modern activation functions, ReLU & Co

Robust loss functions, maximum likelihood and cross-entropy

Weights regularization and dropout

Fifty shades of Deep Learning: ConvNets, pooling, ResNets and BatchNorm

4. Tools of the trade

python, Tensorflow and Keras

Google COLAB and Jupiter Notebook

Deep Learning's “Hello World!”: MNIST

School example in fingerprint localization: UJIndoorLoc

Essential bibliografy