BENDR

Config name: bendr

Pretrained checkpoint: braindecode/braindecode-bendr (HuggingFace Hub)

Reference: Kostas et al., BENDR: Using Transformers and a Contrastive Self-Supervised Learning Task to Learn From Massive Amounts of EEG Data, Front. Hum. Neurosci. 2021

BENDR is a wav2vec 2.0-inspired foundation model for EEG. It consists of a convolutional encoder followed by a transformer contextualizer. The pretrained checkpoint was self-supervised on the Temple University Hospital EEG Corpus (TUEG v1.1/1.2) at 256 Hz using 20 channels (19 standard 10/20 EEG electrodes + 1 relative amplitude channel).

Pretraining data overlap

Warning

The braindecode-bendr checkpoint was pretrained on the Temple University Hospital EEG Corpus (TUEG), which encompasses TUAB and TUEV. These are used both in pretraining and as NeuralBench evaluation datasets, so results on the affected tasks may be inflated.

Pretraining dataset	NeuralBench task	NeuralBench study
TUAB (subset of TUEG)	pathology	Lopez2017
TUEV (subset of TUEG)	clinical_event	Harati2015

Known limitations

• Dropout at inference – The pretrained checkpoint stores drop_prob=0.1; the paper uses 0 during fine-tuning. If needed, this can be overridden via kwargs: {drop_prob: 0.0} in the model config.