NeurIPS 2025 EEG Foundation Challenge: From Cross-Task to Cross-Subject EEG Decoding

Announcements
1

We are pleased to announce the launch of the NeurIPS 2025 EEG Foundation Challenge - From Cross-Task to Cross-Subject EEG Decoding!

banner_square
banner_square1200Ă—1200 187 KB

A major challenge in the deep learning models of EEG recording is the lack of reproducibility. To address this issue, we propose a large-scale machine learning competition!

Here, we are focusing on two things:

  • Advancing the Cross-Task Transfer Learning EEG Decoding:

We need to develop models that can learn across cognitive tasks and can use this knowledge learned.

  • What the models learn needs to be clinically helpful:

Models must learn to generalise across subjects and provide a useful psychopathology clinical measure.

As we are also in the era of the large language model, we have many novelties:

  • Unprecedented Scale and Complexity: We are introducing a large-scale dataset from over 3,000 participants, with high-density 128-channel EEG recordings across six diverse cognitive tasks.

  • Novel Generalisation Scenarios: Challenge 1 focuses on the ability of analytical approaches to generalise to new tasks and new individuals without task-specific adjustments, a crucial step for real-world applicability.

  • Prediction of Psychological Constructs: Challenge 2 uses EEG data to predict latent psychopathology factors (like the p-factor), which could contribute to developing objective markers for mental health research.

This challenge presents a significant opportunity to advance EEG-based decoding and contribute to discovering scientifically validated psychiatric biomarkers.

Are you interested? Stay updated on our website and competition dates.

Let’s brain decode!

Together with:
Bruno Aristimunha, Young Truong, Pierre Guetschel, Seyed Yahya Shirazi,

Isabelle Guyon, Alexandre Franco, Michael Milham, Aviv Dotan,

Scott Makeig, Alexandre Gramfort, Jean-RĂ©mi King, Marie-Constance Corsi,

Pedro A. Valdes-Sosa, Amit Majumdar, Alan Evans, Terry Sejnowski,

Oren Shriki, Sylvain Chevallier, Arnaud Delorme

–

Bruno Aristimunha

PhD Student in Computer Science (Resume)

https://www.linkedin.com/in/bruaristimunha/
Go open-source! bruAristimunha (Bru) · GitHub

2

Hi Bruno and other Authors!
First of all, thank you for this amazing challenge.
I was going through the challenge but it seems so convoluted that I am not able to figure out the 1st task completely. On your website you say,

  • “Participants will train models on passive EEG tasks (Resting State, Surround Suppression, Movie Watching) and evaluate their performance on active tasks (Contrast Change Detection, Sequence Learning, Symbol Search)”

and in your proposal you specifically mention,

  • Section 1.3: “For this challenge, the contestants must predict behavioural performance metrics in an active task (Contrast Change Detection, CCD): response time relative to the start of contrast change and success rate (hit accuracy)”
  • Section 1.4: “Contestants will predict behavioural performance (response time and success rate) in the Contrast Change Detection (CCD) task using EEG data from the Surround Suppression (SuS) task and pre-trial EEG”

So, can you please explain what we have to do?

3

Thanks very much for reaching out.

The proposal is specific about the Challenges, their objectives and how models’ performance will be assessed.

While you are not required to use any of the passive or active task data, except for CCD, to complete Challenge 1, we believe that participant may benefit from pretraining their models on the whole HBN-EEG dataset (w/ or w/o using labels) and then fine tune their models for this specific challenge.

We updated the text on the main page, as well as added more details on the Starter Kit page (Starter Kit | EEG Challenge (2025))

Also, the Google Group is up at https://groups.google.com/u/2/g/neurips2025-eeg-competition. We try to keep track of the question there and respond as soon as possible.

Thanks again,
Yahya

1 Like