Task 2 - Emotion Recognition
In this task participants are asked to infer the emotional category of EEG recordings.
We have 4 emotional labels:
- 0 indicates an emotion with high valence and high power (e.g. Joy)
- 1 indicates an emotion with high valence and low power (e.g. Relief)
- 2 indicates an emotion with low valence and low power (e.g. Sadness)
- 3 indicates an emotion with low valence and high power (e.g. Anger)
Test Dataset.
The test set consists of two parts:
- The held-out-trials test set contains 104 trials, with data coming from the 26 subjects seen during training.
- The held-out-subjects test set contains 128 trials, with data coming from 8 subjects that are not in the training set.
The list of ids to use for inference is in the splits_emotion_recognition.json file.
Target metric
The target metric is the average balanced accuracy score between the two test sets. This means we will compute separately the balanced accuracy score on both held-out-trials and held-out-subjects test sets. These two scores will be averaged together to obtain the final score.
Baseline implementation
We used pruned data as starting point and we applied minimal preprocessing involving:
- deletion of invalid values (nans and outliers)
- standardization with global per-channel statistics
We chose 3 well-known architectures as baselines
- EEGNet [1]
- SyncNet [2]
- EEGChannelNet [3]
No significant changes were applied to the original architectures.
We created two small validation sets:
- val_subjects excluding 2 subjects from the training data
- val_trial extracting 2 trials per subject from the training data.
Training and inference was performed as in task 1 except that we used a majority voting, instead of the average logits.
The final model and hyperparameters (learning rate, batch size) were selected based on the target metric on our valdation sets. A grid search was conducted to optimize these parameters.
Leaderboard
Our strategy yields the following results that serve as baseline
Model | Held-out-trialr Bal. Acc. | Held-out-subject Bal. Acc. | Total Score |
---|---|---|---|
EEGNet [1] | 30.72 | 29.47 | 30.10 |
EEGChannelNet [3] | 29.76 | 22.23 | 26.00 |
SyncNet [2] | 28.80 | 23.08 | 25.94 |
You can replicate results using our GitHub repository.
[1] V j Lawhern et al., “EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces”, J. Neural Eng. 15/5, 2018.
[2] Y. Li et al, “Targeting EEG/LFP Synchrony with Neural Nets”, NeurIPS 2017.
[3] S. Palazzo, C. Spampinato, I. Kavasidis, D. Giordano, J. Schmidt and M. Shah, ““Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features”, in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 11, 2021.