Brant

Brant is a foundation model for intracranial neural signals, developed by the BrainNet group at Zhejiang University (Daoze Zhang, Zhizhang Yuan, and colleagues) and published at NeurIPS 2023. It addresses a long-standing gap in neural signal analysis: while deep learning had transformed scalp EEG and other biosignal domains, intracranial recordings — stereo-electroencephalography (SEEG) and intracranial EEG (iEEG) — lacked a large, reusable pretrained model. Because invasive electrodes are implanted only in clinical settings (primarily for epilepsy monitoring), labeled intracranial data is scarce and task-specific models generalize poorly across patients, electrodes, and downstream objectives.

Brant tackles this with self-supervised pretraining on a large corpus of clinical SEEG data, producing an off-the-shelf encoder that can be adapted to many tasks from a single fixed checkpoint. At the time of release it was described by the authors as the largest model in the brain-signal field, both in parameter count and in the volume of pretraining data.

Brant is the original model in this line of work. Its successor, BrainWave (also called Brant-2), is documented as a separate entry on bio.rodeo and extends the approach to a broader, multi-modal corpus spanning both intracranial and scalp recordings.

#Key Features

• Intracranial-native pretraining: Trained directly on SEEG/iEEG rather than scalp EEG, so its representations are tuned to the high signal fidelity and dense electrode geometry of invasive recordings.
• Joint temporal and spatial modeling: Two stacked Transformer encoders — a temporal encoder and a spatial encoder — capture long-range temporal dependencies within channels and correlations across electrodes.
• Time and frequency information: Inputs combine time-domain and frequency-domain features, letting the model reason about both waveform dynamics and spectral structure relevant to neural activity.
• One checkpoint, many tasks: A single pretrained model supports signal forecasting, frequency-phase forecasting, imputation, and seizure detection without retraining the backbone from scratch.
• Open weights and code: Source code is released on GitHub and pretrained weights are publicly available on HuggingFace under the Apache-2.0 license.

#Technical Details

Brant is a Transformer-based architecture with roughly 500 million parameters, organized as a temporal encoder followed by a spatial encoder so that representations account for dependencies along both the time axis and the electrode array. It is pretrained in a self-supervised manner on 1.01 TB of clinical intracranial neural data recorded via SEEG at a first-class hospital, making it one of the largest pretraining corpora assembled for intracranial signals. The model ingests segmented neural signals enriched with frequency-domain information, enabling it to combine spectral and temporal cues. Evaluated as a fixed, pretrained checkpoint with lightweight task heads, Brant reports state-of-the-art results across neural signal forecasting, frequency-phase forecasting, imputation, and seizure detection, demonstrating that a single pretrained encoder generalizes across distinct downstream objectives.

#Applications

Brant is aimed at clinical and computational neuroscience workflows that rely on invasive recordings. Its most directly clinical task is seizure detection, which supports epilepsy monitoring and surgical evaluation, while its forecasting and imputation capabilities help reconstruct missing or corrupted channels and anticipate signal dynamics in long continuous recordings. Researchers building brain-computer interfaces, neural decoders, or downstream classifiers can use Brant as a pretrained feature extractor, reducing the labeled-data burden that otherwise limits intracranial studies. Because the weights and code are openly available, groups without the resources to assemble a terabyte-scale intracranial dataset can fine-tune or extract embeddings from the released checkpoint.

#Impact

Brant established intracranial neural signals as a viable domain for large-scale foundation models, showing that self-supervised pretraining on terabytes of SEEG data yields transferable representations rather than narrow, task-specific predictors. As an early entry in the brain-signal foundation-model wave, it has been cited in subsequent surveys of brain foundation models and helped motivate later work — including its own successor BrainWave/Brant-2 and other intracranial representation models. Its main limitations stem from the data itself: pretraining draws on clinical epilepsy populations and specific electrode configurations, so generalization to other patient groups, hardware, or recording protocols still warrants validation. The public release of code and weights under a permissive license has nonetheless lowered the barrier for reproducible research on invasive neural recordings.