Machine and Deep Learning
@ INAF-IASF Milano

PROGRAM

• Learning in the context of machine learning
• Different types of machine learning (terminology)
• Main 3 components of learning
• Gradient Descent (the most known optimiser) and its variations
• Typical ML pipeline
• Most used Python libraries in each of the ML pipeline phase
• Most important ML metrics in the case of classification and regression
• One-hot encoding and how it works

• Introduction to python
• Introduction to numpy, pandas, matplotlib
• Minimisation of functions in Python (and the concept of the learning rate)
• Gradient Descent in Python (and its variation: batch, mini-batch, stochastic)
• Non-linear fitting with Python
• scikit-learn library introduction
• Examples of regression and classification with python and scikit-learn

• Unbalanced datasets and what approaches can be used in that scenarios
• PCA and its mathematics (limitations, requirements, etc.)
• Best practices in programming in ML projects (file structure, github repository structure, etc.)
• Model validation and what approaches exist and can be used in which specific scenario
• Neural network introduction
• Keras and its main functionalities

• Model validation with scikit-learn (hold-out approach, k-fold, leave-one-out, etc.)
• Handling of imbalanced datasets with imblearn (oversampling, undersampling, SMOTE, etc.)
• Examples of complete project with astrophysics data
• Example of dimensionality reduction and PCA

• Neural Networks: tasks that can be solved with one neuron
• Convolutional Neural Networks
• Regularisation
• Hyper-parameter Tuning
• Computer Vision

• TensorFlow and Keras (introduction to sequential and functional API, custom training loops, callbacks classes, etc.)
• Neural Networks (feed forward and convolutional neural networks)
• Regularisation (l1, l2, drop-out)
• Example of CNN on a computer vision problem

• Autoencoders
• Generative Adversarial Networks
• PCA
• Error in Labels

• Feature selection (examples)
• Hyperparameter tuning
• Optimisers and neural networks
• Application of advanced topics: callback classes, custom training loops, etc.
• Advanced neural network architectures (autoencoders and generative adversarial networks)

• Dr. Kruk's seminar:
  Exploring astronomy data archives at large scales using deep-learning and crowd-sourcing
• Transfer Learning
• Feature Selection - an introduction