Besides scientfic research, the group aims to upskill the community on AI for Science. The group routinely offers the Machine Learning for Science course to STFC facilities and external institutes, and runs a weekly STFC-wide seminar, inviting eminent speakers from universities, institutes and experimental facilities all over the world.
Machine Learning for Science
This one-week course, designed and developed by SciML, is aimed for delivering both fundamental understanding and practical skills of up-to-date machine learning technologies to domain scientists and students, covering a wide range of topics from the basics of neural networks to some advanced areas such as representative learning and debugging a neural network. The course features practical sessions with real scientific datasets, offering useful suggestions and guidances for applying machine learning to scientific data. The course has been delivered to SCD, Diamond, ISIS and DiRAC.
See our open-source repository for the course at https://github.com/stfc-sciml/sciml-workshop, made of Jupyter Notebooks compatible with Google Colab.
>> Click here to display the Syllabus <<
Classical Machine Learning
| - Supervised regression
- Supervised classification
- Unsupervised clustering
- Practicals with scientific data
|
| Classical ML methods, simple but useful.
|
Deep Neural Networks
| - Basics: network architecture, activation, loss, gradient descent, backprop...
- Introducing TensorFlow
- Practicals with scientific data
| 2
| Open the door to deep learning.
|
Convolutional Neural Network
| - Basics: kernals, pooling, batch normalisation...
- Transfer learning
- Practicals with scientific data
| 2
| Image data deciphered.
|
Autoencoders
| - Basics: dimension reduction, representations, latent space, denoising...
- Supervised, unsupervised, semi-supervised
- Practicals with scientific data
| 2
| Starting point of representative learning.
|
Generative models
| - Theory for generative learning
- Variational autoencoders (VAEs, disentangled VAEs, conditioned VAEs)
- Generative Adversarial Networks
- Practicals with scientific data
| 3
| One of the most intriguing fields in deep learning.
|
Time series
| - Recurrent neural network
- Long short-term memory
- Transformers
- Practicals with scientific data
| 2
| Time series data deciphered.
|
Exploring & debugging
| - Visualising, analysing and understanding neural networks
- Debug neural networks
-
TensorBoard
| 2
| Get your hands dirty.
|
Large-scale resources
| - Using HPC
- Distributed learning on multiple devices
| 2
| Let us gear it up.
|
SciML Seminars
The SciML Seminars are held twice per month. We invite scientists from all over the world to give a 40-minute talk on machine learning, including fundamental developments and applications to solve scientific problems.
- To nominate or volunteer as a speaker, please contact Dr Susmita Basak.
- To receieve information and calendar invitations, please subscribe to our mailist by contacting Dr Susmita Basak, or simply follow us on Twitter.
Photo from www.isis.stfc.ac.uk
