-
Notifications
You must be signed in to change notification settings - Fork 4
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Challenge 22 - XAI for Weather Forecasting Models (Transformer Embeddings) #6
Comments
Hello, Best, Marie |
Hello Marie, I want to team up with you if that's ok with you. Best, Merve |
Hi Marie, Merve and mentors I am interested in this challenge. I'd like to participate. Best, |
Hi Marie, |
Hi Marie, I'm also interested in this challenge. And I have been working on XAI for my project recently. May I team up with you? Best, |
Hello, |
I received your invitation to a repository, thank you! |
Oh, yep. Should we have a messenger group or something like that? |
Hello, I created a discussion channel in the repository, we can start with this to share ideas. But I think you're right it would be good if we could schedule an online meeting in the coming days. |
Hi, I am also interested in joining the team if still taking applications :). My background is in data science/engineering and geophysical fluid dynamics. Best, James |
Hi James, yes, of course, let me add you to the repository Best, |
Hello @mariebrl, |
Hello @mc4117, |
Hi Scarlet @scarletto999, GIven the time limitations, we think it would probably be better to use a pre-trained transformer and explain post-hoc |
Hello @mc4117, |
Hi Marie, You should find more about the model architecture here https://www.ecmwf.int/en/about/media-centre/aifs-blog/2024/first-update-aifs Mariana |
Hello Mentors, we wonder if you expect us to propose a general approach suitable for "weather transformers" or if we should be more specific about an approach to explain PanguWeather? There is no documentation yet on AIFS and thus it is impossible to come up with a method tailored to its architecture. Scarlet |
Hi, |
Hi, Prashant |
Hello, chicha |
Hello, I would like to participate in this challenge. May I ask, what are the expected deliverables for this challenge? Like code, reports, etc..? Best regards, |
Hi, I'm also interested in this challenge. I have experience on NWP, data processing, validating but would like to explore XAI weather models. There might some people who could team up in a different team if there is already too many people on @scarletto999 @sandupal @kam3545 @chicha1986 @thinhngo-x . Just show you interest and we can create a repo to work on. |
Hello in terms of deliverables, for each of the challenges, teams are expected to prepare a presentation to present their work and findings as well as to share the code in github |
Dear weather team, Regards, Teklehaimanot |
Hi everyone here, kind regards, |
Dear @Teklehaim, If you want to participate in Code for Earth, you need to submit a proposal, explaining how you would solve the stated problem of the challenge. The deadline for submission ends 9. April and then we will evaluate the proposals. Those selected will work over the summer on their projects and the mentors will provide some guidance. For more information, please read the FAQ and Terms & Conditions on our website. Additionally you might want to listen to the recording of the Q&A session on Youtube. I hope this gets you going! Bye, Athina
|
Thank you very much @trakasa for the information. I sent my application today. Regards, |
Hi, I was checking the apply form and it does not work. It should until today end of the day, right? Thank you |
Dear @oriollacave Please let us know, if you still have any issues. Apologies for the inconvenience and thanks for your patience! Bye, Athina
|
@oriollacave ... forgot to mention: yes, deadline for submission is tonight at 23:59 CET ;-) |
Challenge 22 - XAI for Weather Forecasting Models (Transformer Embeddings)
Goal
Welcome to the XAI Transformer Embedding Challenge for Weather Forecasting Models! In this challenge, participants will explore the explainability of Transformer embeddings in the context of weather forecasting models, using models like PanguWeather. The goal is to develop insights and techniques that enhance the interpretability of these AI models. Additionally, participants are encouraged to consider the possibility of utilizing visualization techniques similar to the ones developed by BertViz, Quantus or SHAP tailored to weather forecasting data.
Mentors and skills
Challenge description
Traditional weather forecasting models often rely on solving physical equations to make a forecast. However, the emergence of deep learning and transformer-based models has shown potential in improving forecast accuracy. Transformers, renowned for their effectiveness in natural language processing tasks, have been adapted to time series forecasting, including weather prediction due to their ability to capture complex spatiotemporal patterns in weather data. However, as these models become more complex, understanding how they arrive at their predictions becomes crucial for trust, accountability, and further model improvements.
Ideas for the challenge implementation:
The challenge involves exploring and analyzing the weather data used to train the AI weather forecasting models to gain insights into the patterns, trends, and relationships between different weather variables.
Participants will leverage pre-trained Transformer models (such as PanguWeather, AIFS* etc.) fine-tuned on weather forecasting tasks to analyze the embeddings generated by these models and understand how they encode information about weather patterns and features.
Participants will develop explainability techniques tailored to Transformer embeddings in the context of weather forecasting, to enhance model interpretability, providing insights into how specific weather events or patterns are represented in the embeddings and how they contribute to the final predictions.
Participants will evaluate the effectiveness of their explainability techniques using relevant metrics. They will interpret the insights gained from their techniques and assess their utility in improving the interpretability and trustworthiness of weather forecasting models.
Participants will also be encouraged to implement visualization techniques tailored to weather forecasting data, to visualize attention mechanisms and model predictions (inspired by BertViz).
The text was updated successfully, but these errors were encountered: