Rules | EEG Challenge (2025)

The rules provided to contestants are as follows:

Evaluation and testing will be done on the downsampled data. To ensure timely evaluation, we will downsample the data to 100 Hz after filtering the data to the 0.5-50 Hz range. The down sampled datasets will be used for evaluation and testing. The downsmapling script is available on the downsample-datasets repository.
Contestants are allowed and encouraged to use any datasets for pre-training. However, they must clearly mention and document any additional data sources used in their submission.
Contestants are allowed to use existing foundation models, but they must clearly mention and document which pre-trained models were used and how they were fine-tuned in their submission.
Contestants must submit their code during the inference stage; this is a code submission competition.
Models must be able to run on a single GPU with 20 GB of memory at the inference stage. Contestants are encouraged to validate their models on a similar machine before submitting their code.
Related members from the organizing team can participate but are ineligible for prizes.
The top 10 teams will have their code released after the final submission.

Evaluation Criteria

The metric used for challenge 1 is the normalized root mean square error for the response time prediction.
The response time should be quantified for each trial.
We decided not to use correct/incorrect classifications for this challenge.

Formally, the score will be computed as follows:

from numpy import std
from sklearn.metrics import root_mean_squared_error as rmse
score = rmse(y_trues, y_preds) / std(y_trues)

Similarly, challenge 2 also uses the normalized root mean square error.
We encourage participants to use all tasks to make the inference.
At the inference stage, we will only use the XXX (will be decided soon after more testing) task to predict the metrics. This is to ensure a timely evaluation due to the competition scale.