LaBraM++

LaBraM++ is a large brainwave foundation model that rebuilds the vector-quantized tokenizer used by LaBraM around a correction from signal processing theory. It comes from an independent team — Konstantinos Barmpas, Na Lee, Yannis Panagakis, Dimitrios Adamos, Nikolaos Laskaris, and Stefanos Zafeiriou at Imperial College London, the BCI company Cogitat, the National and Kapodistrian University of Athens, and Aristotle University of Thessaloniki — not from LaBraM's original authors, and was presented at the NeurIPS 2025 Foundation Models for the Brain and Body workshop. The "++" marks an architectural critique and reimplementation by outside researchers, not an official successor.

The critique is specific and mathematical. LaBraM's tokenizer learns its neural codebook by reconstructing the Fourier amplitude and phase of one-second EEG patches, penalizing phase error with a squared difference between angles. Phase is periodic on [-π, π], so that loss is discontinuous at the wrap-around: a predicted phase of 179° against a true phase of -179° — nearly identical physiological states — incurs a loss approaching 4π² and a gradient approaching 4π, far larger than anywhere else in the domain. The result is unstable optimization and a representation bottleneck at the phase values where oscillatory timing information lives.

LaBraM++ replaces that term with separate sine and cosine reconstruction losses, which together equal 2 − 2cos(ϕ̂ − ϕ) — proportional to the squared chord length between two points on the unit circle. The 179°/-179° case now costs almost nothing, gradients stay bounded across the domain, and the tokenizer interpolates smoothly through the wrap.

#Key Features

• Circular phase representation: Phase is reconstructed as a point on the unit circle via paired sine and cosine targets, making reconstruction error geometrically meaningful and removing the gradient discontinuity at ±π.
• Flexible sample construction: Every pretraining sample is fixed at 256 patches of 200 time points, with channels and time window drawn randomly from an indexed recording, so the per-dataset window-count tuning the original design required disappears.
• Global electrode indexing: Spatial embeddings index a single 104-electrode list covering the whole pretraining corpus, so heterogeneous montages share one embedding table.
• Signal-conditioning ablations: Common average referencing, per-patch Z-scoring, and right-aligned temporal embeddings are optional additions; the combination gives the best mean score across tasks.
• Improved optimization: Under identical backbone, data, and hyperparameters, the redesigned tokenizer yields a substantially lower pretraining loss.

#Technical Details

The tokenizer keeps LaBraM's overall shape — a temporal convolution stem, neural transformer encoder, an 8192-entry codebook of 64-dimensional tokens, and a shallow reconstruction decoder — and changes the objective to amplitude plus sine-phase plus cosine-phase plus the quantization loss. It was trained for 100 epochs; the foundation model, a LaBraM-base-sized 12-layer transformer with hidden dimension 200, 10 attention heads, and MLP dimension 800, was pretrained for 50 epochs on eight NVIDIA L4 GPUs. Pretraining used the publicly available datasets from LaBraM — including BCI Competition IV-1, Grasp and Lift, PhysioNet MI, Siena Scalp, TUAR, TUEP, and TUSZ — plus roughly 235 hours of the authors' own motor-imagery recordings standing in for LaBraM's private corpus, all bandpass filtered 0.5–44.5 Hz and resampled to 200 Hz.

Evaluation uses a four-task benchmark (motor, memory, sleep, eyes-open/closed) with 10-fold subject-independent cross-validation and full fine-tuning for 20 epochs. Both models were re-pretrained from scratch on the same corpus for a controlled comparison: LaBraM++ reaches 0.722 mean balanced accuracy against 0.664 for LaBraM, with the largest gain on the motor task (0.723 versus 0.570). The conditioning variants raise the mean to 0.726. Against other open-source brainwave models on the same protocol, LaBraM++ leads NeuroGPT's encoder variant (0.713), CBraMod (0.666), and EEGPT (0.566). Reconstruction remains imperfect: the codebook captures overall trend and low-frequency structure but loses higher-frequency detail, which the authors propose addressing with scattering transforms.

#Applications

The model targets the brain-computer interface and clinical EEG workloads its predecessor addressed — motor-imagery decoding, memory-task classification, sleep staging, and eyes-open/closed detection — where labeled recordings are scarce and unlabeled ones plentiful. Montage-agnostic input handling lets one pretrained backbone be fine-tuned across hardware with different electrode counts and placements, which matters for consumer headsets and for pooling clinical archives recorded under different protocols.

#Impact

LaBraM++ is a corrective in a subfield that has largely borrowed architectures from vision and language without auditing whether their loss functions respect the physics of the signal. Its central result — that a small change to a phase loss buys several points of balanced accuracy — argues that headroom in EEG foundation models lies in signal-aware design as much as in scale. The evidence base is correspondingly modest: a workshop paper covering four downstream tasks with a single base-sized backbone, whose LaBraM comparison uses a re-pretrained baseline rather than the published checkpoint. The MIT-licensed repository accompanying the paper currently contains only a README and the license file, with no code or weights released.