Upcoming talks
Thursday 19 October
Speaker: Magnus Ørke
Title: Deep learning algorithms for solving high-dimensional PDEs
Abstract: Deep learning approaches to solving PDEs have recently gained some attention. I will give an overview of the most promising methods and their results, specifically for high-dimensional problems and the attempt to overcome the curse of dimensionality, and discuss challenges and possibilities related to these methods.
Past talks
Thursday 28 September
Speaker: Adrian Ruf
Topic: Error bounds for physics informed neural networks
Thursday 21 September
Speaker: Johan Wind
Topic: (Will continue with PINNs from last time)
Thursday 14 September
Speakers: Vegard Antun, Johan Wind
Title: Introduction to PyTorch (continued)
Abstract: In the first half of the seminar, Vegard will finish his introduction to PyTorch from last time.
Then, Johan will show how to solve Laplace's equation with a simple PINN (Physics Informed Neural Network). Johan invites people to code along, given that you have a laptop with PyTorch set up. For example, you may use Jupyter Notebooks like Vegard showed last time (https://github.com/vegarant/mievu4020_v23/blob/main/Working_remotely.md).
Thursday 7 September
Speaker: Vegard Antun
Abstract: In this talk, I will give a brief introduction (on the blackboard) to neural networks and the standard setup for training these networks in a supervised fashion. Then, I will look at how one can perform this type of training on a computer using PyTorch (one of the dominant machine learning frameworks for training neural networks in Python). Specifically, I'm planning to cover:
- How to run PyTorch on the university's computers, and how one can work remotely on these computers from your own laptop.
- Tensors and basic mathematical operations on tensors.
- Loading data and splitting it into batches.
- How one can compute gradients of functions in PyTorch.
- Different ways of creating neural networks and training them in PyTorch.
- If time allows, I will also cover some bits of the functional API.
The talk will be based on some of the teaching material I developed for the course MIEVU4020. See this page for details.