GradDays are giving courses that broaden the physics knowledge of students and teach specialized useful techniques.
Program of GradDays included:
- Black Holes and Quantum Gravity Constraints on Field Theories
- Correlated Quantum Dynamics of Ultracold Few- to Many-Body Systems
- Quantum Simulation with Quantum Optical Systems
- The formation of structure in cosmology
- Solar System Exploration Missions and their Scientific Outcome
- Quantum Field Theory in Extreme Environments
as well as other sweet things.
Even so, our course “Machine learning and applications in Science and Industry” was the most popular. Focus of the course (heavily influenced by time constraints: only 4 days) was to give a wide overview of useful models in Machine Learning and their applications in very different areas, and even contained optional practice!
That’s why we put inside many interactive demonstrations of machine learning techniques!
Also we tried to create a nice bridge between models and their real-life applications. Many of the examples were from particle physics — an area that we’re working in (tracking, tagging, reweighting, uniform boosting, particle identification, simulation refinement, tuning of simulation parameters, etc.). However we also included some notable examples from other data-intensive areas: astronomy, neuroscience, medicine, climatology and biology.
Finally, many other interesting things done with machine learning were discussed: spam detection, search engines, visual recognition, kinect and AlphaGo, recommender systems and news clustering.
Lecture of the first day gives some introduction into problems, applications and notions of machine learning. Several simple models are discussed to get an impression:
- knn and search for neighbours
- density estimation techniques
- mixtures of distributions
- clustering methods
- linear models with regularization.
In the second day we made focus on tree-based techniques, specially boosting, that aren’t popular in research now, but work very well in practice and are best-performers in many examples with tabular data
- decision trees for classification and regression
- Random Forest
- AdaBoost and Reweighter
- Gradient Boosting for classification, regression and ranking (ordering of items)
- Uniform boosting
- applications: particle identification, triggers and search engines
On the third day we get back to continuous optimization models, start from revisiting linear and generalized linear models, then more involved models are introduced
- linear models and their generalizations
- regularizations (again)
- SVM and kernel trick
- spam detection and elements of visual recognition
- factorization models and recommender systems
- factorization machines
- unsupervised dimensionality reduction techniques (PCA, LLE and IsoMAP)
- supervised dimensionality reduction techniques: CSP and LDA
- artificial neural networks
Finally, the last day was devoted mostly to deep learning: convolutional and recurrent neural networks, autoencoders, embeddings, GANs and others.
Also, an active learning was demonstrated in couple with gaussian processes.
- interactive demonstrations of machine learning models
- thanks to authors of all materials / demonstrations that were used in the course
- materials of GradDays at github (including practice assignments in jupyter notebooks)