iBrain

An electrode grid on the cortical surface and a microelectrode array pushed into motor cortex are listening to the same tissue, a few millimetres apart. But one delivers a continuous voltage waveform sampled at hundreds of hertz, and the other a sparse count of action potentials per millisecond bin. Those two data types are different enough that pretrained models for invasive neural recordings have been built separately for each, so each already-scarce corpus trains its own model.

iBrain is a foundation model that pretrains on both at once. It keeps the front end signal-specific — a convolutional adapter reads intracranial EEG (iEEG) waveform patches, an MLP reads binned spike counts — and then shares everything downstream: a single spatiotemporal Transformer contextualizes the resulting tokens across recording channels and across time, regardless of which kind of electrode produced them. The wager is that surface field potentials and single-unit spiking are complementary views of the same dynamics, and that a shared channel–time inductive bias is worth more than modality-pure representations.

The model comes from the Chinese Institute for Brain Research and was posted as a preprint in September 2026. It joins two established lines of work: Brant and BrainBERT pretrain on intracranial field recordings, while the Neural Data Transformer family — most recently NDT3 — and POYO pretrain on intracortical spiking. Unified pretraining across heterogeneous brain signals has mostly been attempted on non-invasive modalities, as in BrainOmni's joint EEG–MEG encoder; iBrain applies it to invasive recordings only.

#Key Features

• Signal-specific tokenization: One-second segments split into ten 100 ms patches. For iEEG at 500 Hz each patch holds 50 waveform samples and passes through a temporal convolutional adapter with a residual linear pathway; spikes are binned at 20 ms, giving five counts per patch through an MLP adapter.
• Criss-cross spatiotemporal attention: Each block applies channel attention independently at every time index, then temporal attention along each channel, with a learnable relative temporal bias and a signal-type embedding.
• Losses matched to the signal statistics: Masked reconstruction uses squared error on channel-normalized iEEG waveforms and a Poisson negative log-likelihood on softplus-rectified rates for spike counts, scored only at masked, non-padded positions.
• Channel-view alignment: Two overlapping subsets of the valid channels are encoded separately and their pooled representations aligned with a SimSiam-style stop-gradient cosine objective, so representations survive electrode subsampling.
• Frozen-backbone transfer: Used as a fixed feature extractor rather than fine-tuned, the encoder still reaches R² 0.904 on MC-Maze.

#Technical Details

The backbone is a 6-layer Transformer with hidden dimension 256, 8 attention heads, a feed-forward dimension of 1,024 and 0.1 dropout — small by foundation-model standards. Pretraining masks 50% of valid channel–time tokens, samples two channel views retaining roughly 80% of channels each with 75% overlap, and alternates iEEG and spike minibatches at a 1:1 step ratio for 30 epochs of AdamW on 8 GPUs. The corpus is 3,950 hours of human ECoG and sEEG from AJILE12 and the SWEC iEEG Dataset plus 3,209 hours of spiking from the Neural Pile — about 7,160 hours and 25.8 million one-second samples.

On four Brain Treebank iEEG tasks — audio pitch, volume, sentence onset and speech detection — fine-tuned iBrain reaches AUCs of 0.68, 0.89, 0.86 and 0.89 against 0.60, 0.75, 0.81 and 0.79 for the strongest baseline on each. On spike decoding it reports R² of 0.914 on MC-Maze, 0.903 on Area2-Bump and 0.785 on Perich T-CO, above NDT1, NDT2, MtM, POYO and UniBCI, but 0.692 on Perich T-RT, where UniBCI's 0.716 is better — seven of the eight benchmarks. A scaling sweep from 500 to 7,160 hours lifts MC-Maze R² only from 0.882 to 0.914, with gains flattening past 2,000 hours, which the authors read as redundancy within subjects and sessions.

#Applications

The immediate users are groups running invasive brain–computer interfaces, where every hour of recording costs a surgical implant and participant time. A pretrained initialization that recovers most of its full-data accuracy from half the labeled trials — on Area2-Bump, R² climbs from 0.588 at 10% of labels to 0.755 at 20% — shortens calibration on new sessions and implants, and one checkpoint covers motor decoding from intracortical arrays alongside language-related responses from clinical iEEG montages.

#Impact

iBrain's contribution is the demonstration that joint pretraining across two invasive recording modalities beats pretraining on either alone, on both signal types — an argument for pooling the fragmented datasets that individual labs and epilepsy monitoring units accumulate. Its own scaling result tempers that: recording hours alone buy little past 2,000, and diversity across subjects, tasks and acquisition setups is the variable that matters. It is a preprint awaiting peer review, no code or weights have been released, and the evaluation pairs human iEEG with macaque spiking rather than both signals in one subject.