The syllabus of the course is available here. The repo also contains the exercises and solutions of the course. The course is based on my Physics Report Data science applications to string theory. If you find the code provided here useful in your research, please consider citing it. Videos of the lectures can be found here (requires access to Oxford's weblearn contents).
In this exercise, we set up a Python environment for Machine learning. We create our first feed-forward neural network that learns to classify data into two or three classes using Keras and PyTorch.
We also use CMS collider data, publicly available from CERN opendata, to learn to predict the invariant mass of the Z boson in a Z → e+ e- decay. The repository for the exercise with templates and solutions is available here.
In this exercise, we create our first convolutional neural network using PyTorch. We classify galaxies into spiral, elliptical or unknown. The data is provided by the Galaxy Zoo project. See this publication for more details. The pictures of the galaxies themselves are provided by the Sloan Digital Sky Survey. The repository for the exercise with templates and solutions is available here.
NN preidction:
 |  |  |
| spiral: 82% elliptical: 10% unknown: 8% | spiral: 8% elliptical: 90% unknown: 2% | spiral: 12% elliptical: 10% unknown: 78% |
(Image source: https://www.sdss.org)
In this exercise, we demonstrate how to code an environment that can be connected via the OpenAI gym to ChainerRL. We illustrate how the the A3C agent finds good energy configurations for the 1D Ising model by flipping spins at any of the lattice sites. The repository for the exercise with templates and solutions is available here.
I found an optimal configuration!
↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑
I started from
↑ ↑ ↓ ↓ ↑ ↑ ↑ ↑ ↑ ↓ ↓ ↑ ↓ ↑ ↑ ↓ ↑
and took the actions
[10, 2, 3, 9, 12, 15]
In this exercise, we illustrate different unsupervised clustering algorithms (k-means, mean shift, DBSCAN, Birch) that were discussed in class using scikit learn.
We furthermore use genetic algorithms to minimize a superpotential of an N=1 SUSY theory, which arises e.g. from string theory.
The repository for the exercise with templates and solutions is available here.