ARTIFICIAL INTELLIGENCE FOR ASTROPHYSICAL PROBLEMS
- Overview
- Assessment methods
- Learning objectives
- Contents
- Full programme
- Bibliography
- Teaching methods
- Contacts/Info
Statistics and elementary concepts of programming in Python
The final examination is composed by three steps:
1. Implementation of a full end-to-end ML project assigned during the lesson in which the student is invited to implement and test ML algorithms to real (or simulated) dataset given by the teacher.
2. A report that describes the applied ML algorithms and discusses critically the results, especially in the domain of Astrophysics.
3. An oral examination that covers the topics of this Course.
Point (1) and (2) are weighted to 70% of the final score, while point (3) account for the remaining 50%. The final score will be calculated by weighting the scores obtained in each pint. The goal of the examination is twice:
- Check that the student has correctly understood the formal framework behind ML, Deep Learning and Big Data
- Check that those concepts are fully absorbed when applied to real astrophysical problems.
The Machine Learning (ML) and Deep Learning is a branch of Applied Statistic that has demonstrated to be a breakthrough in many different areas and science that are driven-by-the-data- Astrophysics is not an exception and in the near future the huge availability of data from various survey (e.g. LSST, CTA, SKA) will require the adoption of state-of-the-art BigData paradigm and ML algorithm for doing science.
At the end of the course, the student will be able to:
- Understand the main concepts of Machine Learning and Deep Learning with a solid background on their foundations, required for application of those tools to real-life problems in Astrophysics.
- Apply, understand end extend classification algorithms (Bayesian Classifier, Random Forest, Multi-linear classifiers) exploring their capabilities to the automatic classification of astrophysical transients.
- Explore Deep Learning Neural Networks for image processing applying them to Time Domain Astrophysics (discover of transients). At the end, the student will be able to handle concepts related to Neural Network, Deep Learning and validation of ML algorithms.
- Use the state-of-the-art Cloud Platform and BigData suite from Google to simulated LSST data.
1) Introduction to Machine Learning techniques
1a) Training Set, Test Set, Validation, Scores, ROC curves
2) Classification algorithms. Case I - transients classification in Transient Astrophysics. The LSST survey
2a) Basic concepts, why we need machine learning
2b) Likelihood based classifiers, Bayesian Classifiers, Naive Bayes classifier, K-Nearest Neighbors, Random Forests and their optimization, Linear classifiers
2c) Metrics and optimization, K-Fold validation. Concepts of hyperparameter optimization
3) Image recognition algorithms - Case II - Discover of transients in Time Domain Astronomy with CCD images
3a) Neural Network concepts
3b) Training and algorithms
3c) Deep learning - image classification
4) Elements of unsupervised algorithms in Machine learning (e.g K-Means, clustering techniques).
5) State of the art with Machine Learning in Big Data era
5a) The case of Google Cloud Platform. Application of the platform to LSST mock up data.
6) Hands on session.
-
- Pattern Recognition and Machine Learning (C. Bishop – Springer Edition)
- Deep Learning (Adaptive Computation and Machine Learning series) – Ian Goodfellow ISBN-13: 978-0262035613
- Deep Learning – John D. Kelleher - The MIT press.
The course is organized in regular frontal lectures with formal introductions to the covered topics followed by their application to real problems. Some hands-on sessions are foreseen in order to allow students to test and practice with algorithms and framework discussed during the frontal lectures.
For any question, discussion, concern, etc, students are invited to contact the teacher at the following email: marco.landoni@inaf.it