Uploaded June 2022 | Updated November 2026, 3 days ago
Jesper Dramsch is a scientist for machine learning at the European Centre for Medium-Range Weather forecasts. They have a phd in applied Machine Learning to Geoscience from Technical University of Denmark. They are a Kaggle Kernals Expert and TPU star, ranking at top 81/100k worldwide. We talked about weather forecasting, things they learned from Kaggle, how to deal with missing data and ourliers, deep learning, Keras vs Pytorch, XGBoost, their struggles as a phd student, working in the EU vs US. If you like the show subscribe to the channel and give us a 5-star review. Subscribe to Daliana's newsletter on dalianaliu.com for more on data science.
Daliana's LinkedIn: linkedin.com/in/dalianaliu
Daliana's Twitter: twitter.com/DalianaLiu
Resources shared by Jesper:
The newsletter with missing data:
https://buttondown.email/jesper/archive/towels-have-quite-a-dry-sense-of-humor/
The paper by Gael about missing data:
academic.oup.com/gigascience/article/doi/10.1093/gigascience/giac013/6568998
The Huber Loss:
en.wikipedia.org/wiki/Huber_loss
Skill Scores:
en.wikipedia.org/wiki/Forecast_skill
Brier Skill in Weather:
dwd.de/EN/ourservices/seasonals_forecasts/forecast_reliability.html
CRPS Continuous Ranked Probability Score
datascience.stackexchange.com/questions/63919/what-is-continuous-ranked-probability-score-crps
ConvNext, Convnets for the 2020s:
arxiv.org/abs/2201.03545
Transformers for ensemble forecasts:
arxiv.org/abs/2106.13924
Books recommend:
amazon.com/shop/jesperdramsch/list/2DYS5KVR5TX0E
Blog posts I wrote about these books:
dramsch.net/tags/books
Short I made about Test-Time Augmentation
youtube.com/shorts/w4sAh9lKyls
Their links: dramsch.net/links
Their open PhD thesis: dramsch.net/phd
Newsletter: dramsch.net/newsletter
Twitter: dramsch.net/twitter
Youtube: dramsch.net/youtube
Linkedin: dramsch.net/linkedin
Kaggle: dramsch.net
0:00 Intro
00:01:27 how he got into in ML
00:07:22 choose models depending on the type of data
00:09:10 how he handled missing data
00:23:49 how he validated the model
00:28:34 Transformers are eating the world
00:31:15 Deep Learning for small dataset
00:34:16 never underestimate the domain experts
00:49:36 Hoover Loss is a fantastic metric to deal with extreme values
00:54:48 his experience with Kaggle competition
01:02:59 Kaggle tricks that helped his models perform better
01:08:18 PyTorch vs Keras
01:21:11 how he managed the stress
01:30:30 working in different countries and cultures
01:42:52 the books and blogs he recommended about ML
Jesper Dramsch is a scientist for machine learning at the European Centre for Medium-Range Weather forecasts. They have a phd in applied Machine Learning to Geoscience from Technical University of Denmark. They are a Kaggle Kernals Expert and TPU star, ranking at top 81/100k worldwide. We talked about weather forecasting, things they learned from Kaggle, how to deal with missing data and ourliers, deep learning, Keras vs Pytorch, XGBoost, their struggles as a phd student, working in the EU vs US. If you like the show subscribe to the channel and give us a 5-star review. Subscribe to Daliana's newsletter on dalianaliu.com for more on data science.
Daliana's LinkedIn: linkedin.com/in/dalianaliu
Daliana's Twitter: twitter.com/DalianaLiu
Resources shared by Jesper:
The newsletter with missing data:
https://buttondown.email/jesper/archive/towels-have-quite-a-dry-sense-of-humor/
The paper by Gael about missing data:
academic.oup.com/gigascience/article/doi/10.1093/gigascience/giac013/6568998
The Huber Loss:
en.wikipedia.org/wiki/Huber_loss
Skill Scores:
en.wikipedia.org/wiki/Forecast_skill
Brier Skill in Weather:
dwd.de/EN/ourservices/seasonals_forecasts/forecast_reliability.html
CRPS Continuous Ranked Probability Score
datascience.stackexchange.com/questions/63919/what-is-continuous-ranked-probability-score-crps
ConvNext, Convnets for the 2020s:
arxiv.org/abs/2201.03545
Transformers for ensemble forecasts:
arxiv.org/abs/2106.13924
Books recommend:
amazon.com/shop/jesperdramsch/list/2DYS5KVR5TX0E
Blog posts I wrote about these books:
dramsch.net/tags/books
Short I made about Test-Time Augmentation
youtube.com/shorts/w4sAh9lKyls
Their links: dramsch.net/links
Their open PhD thesis: dramsch.net/phd
Newsletter: dramsch.net/newsletter
Twitter: dramsch.net/twitter
Youtube: dramsch.net/youtube
Linkedin: dramsch.net/linkedin
Kaggle: dramsch.net
0:00 Intro
00:01:27 how he got into in ML
00:07:22 choose models depending on the type of data
00:09:10 how he handled missing data
00:23:49 how he validated the model
00:28:34 Transformers are eating the world
00:31:15 Deep Learning for small dataset
00:34:16 never underestimate the domain experts
00:49:36 Hoover Loss is a fantastic metric to deal with extreme values
00:54:48 his experience with Kaggle competition
01:02:59 Kaggle tricks that helped his models perform better
01:08:18 PyTorch vs Keras
01:21:11 how he managed the stress
01:30:30 working in different countries and cultures
01:42:52 the books and blogs he recommended about ML










