Introduction to Uncertainty Quantification https://events.hifis.net/event/2767/ Please share with us, why you are interested in Uncertainty Quantification! - PS: I like error bars - ... -I am curious to know what exactly is uncertainty quantification, I don't have any prior knowledge on the topic, but I heard about it in my faculty corridors. :) - I had it mention in a course during my Masters and throught it to be quite important but often overlooked - I like understanding precision and accuracy <3 - and hope to have a job as data scientist in a near future >D - I want to be able to assess the quality of the model created and say how good or bad it is -It is essential for combining multi-source data (reanalysis, in-situ, remote sensing etc.) - I need UQ for my project in near future - Because I like to explore the uncertainty in life - I do research in this area and want to see if I am missing out on topics - I want to understand how reliable my models and simulations (in physics we rarely deal with perfect information, everything is noisy and approximate) - I want to obtain new methods to cope with the uncertainty in real-world data - Check to what degree statistical variance estimation does (not) generalize to ML methods - Determine model quality in terms of generalization - I want to have a better overview about the quality of outcomes of data driven methods Please share with us, what machine learning task is driving your curiosity about uncertainty quantification! - PS: LLMs - RF, GX, deep learning - ... interseted in Rondom Forest , deep learning, XGBoost, time series analysis and ML methods like LSTM - PCA-SVM Classification + CNN - Genertive AI, espeically synthetic data generation algorithms - Autonomous driving, deep learning-based perception / predictions modules - Learning it for future projects currently image reconstruction - Reinforcement Learning, OOD generalisation -Deep learning tasks in general - I am interested in data analysis and prediction - My postdoc project on magnetic field simulationsand forward–inverse problems:) And it includes a lot of subtasks - SS: essential for ML - Classification & regression tasks -Blending observation, remote sensing and reanalysis through ML models that can output condidence intervals For those who would like to get free support on their ML project, please consider approaching us Helmholtz AI consultants for this. More details can be found here: https://www.helmholtz.ai/you-helmholtz-ai/ai-consulting/ Want to stay connected? - upcoming workshop: https://events.hifis.net/event/3142/ - mattermost channel: https://mattermost.hzdr.de/signup_user_complete/?id=kia8w7m7ejd8bxjmpr7xf61h5w&md=link&sbr=su Share something that you liked about the course or something that you learned! - ... - Very nice content! Using notebooks to explore and get a feel for what happens is a great way to learn! - I liked the coverage of the course. Starting with basics and going up to recent research examples. - I learned what uncertainties are and their capital role to guarantee the reliability in our foundings, some techniques to compute uncertainties. I liked the fact that the notebook where nicely commented, so th at I didn't feel very lost- - I liked the part about the calibration plot and how to evaluate the results based on that - I love notebooks, definitely will take a look closer on all this data later - Very nice general overview Share something that you wish could be improved! - ... - Going through notebooks and analyzing their results (I understand that time of course is short but maybe extend the time of course) - Clarify what disributional vs deterministic uncertainty quantifier are or briefly that this distinction exist. - nice workshop! the coding time is a bit too long, maybe can consider student take the coding home. In the course more focus on the coding results interpretation. - Maybe explain the results a little bit more - The workshop (understandably) focusses on UQ given a certain (type / specification / fitting strategy) of models. But sample selection itself also causes a lot of uncertainty. Maybe give at least a few hints on that - I think the "lecture" part could be sped up a bit. I would also have liked to see some "take home messages" for the notebooks (i.e. what works in the scenarios and why.) -I feel that there is a need of more homework before the lecture itself/divide it by two parts and have lecture on 1 day and more practice on another - more self-study sources