Data Science Quick Reference Manual – Deep Learning

Book excerpt: This work follows the 2021 curriculum of the Association for Computing Machinery for specialists in Data Sciences, with the aim of producing a manual that collects notions in a simplified form, facilitating a personal training path starting from specialized skills in Computer Science or Mathematics or Statistics. It has a bibliography with links to quality material but freely usable for your own training and contextual practical exercises. Part in a series of texts, it first summarizes the standard CRISP DM working methodology used in this work and in Data Science projects. As this text uses Orange for the application aspects, it describes its installation and widgets. The data modeling phase is considered from the perspective of machine learning by summarizing machine learning types, model types, problem types, and algorithm types. Deep Learning techniques are described considering the architectures of the Perceptron, Neocognitron, the neuron with Backpropagation and the activation functions, the Feed Forward Networks, the Autoencoders, the recurrent networks and the LSTM and GRU, the Transformer Neural Networks, the Convolutional Neural Networks and Generative Adversarial Networks and analyzed the building blocks. Regularization techniques (Dropout, Early stopping and others), visual design and simulation techniques and tools, the most used algorithms and the best known architectures (LeNet, VGGnet, ResNet, Inception and others) are considered, closing with a set of practical tips and tricks. The exercises are described with Orange and Python using the Keras/Tensorflow library. The text is accompanied by supporting material and it is possible to download the examples and the test data.