Device Agnosticism

What It Means

A device-agnostic Brain Foundation Model generalises across EEG hardware generations and manufacturers. The same pre-trained encoder processes signals from a 500 Hz research-grade Neuroscan amplifier and a 256 Hz consumer Muse S headband without any hardware-specific preprocessing branches or separate model weights.

Why Hardware Matters

EEG amplifiers differ along several dimensions that affect the recorded signal:

These differences manifest as systematic biases in the recorded signal that a naive classifier will learn as confounders - effectively memorising which amplifier was used rather than which cognitive state is occurring.

Approaches to Device Agnosticism

Multi-Source Training

The most direct approach is to pre-train on recordings from as many different hardware platforms as possible. When the training distribution spans diverse hardware, the encoder learns features that are shared across hardware (i.e. true neural signal features) rather than hardware-specific artifacts.

Robust Normalisation

Per-channel z-score normalisation or robust normalisation (median absolute deviation) within each window removes the DC offset and amplitude scale differences between hardware, making the representation invariant to absolute amplitude.

Riemannian geometry-based covariance representations are inherently scale-invariant: a covariance matrix normalised to unit trace is unaffected by uniform amplitude scaling, making it a hardware-agnostic feature representation.

Augmentation-Based Domain Randomisation

During pre-training, artificial hardware-like degradation is applied as data augmentation:

• Adding correlated pink noise to simulate amplifier noise floors.
• Random amplitude scaling to simulate different gain settings.
• Baseline drift injection to simulate poor electrode contact.
• Additive 50/60 Hz sinusoidal noise to simulate power-line interference.

By training the model to be invariant to these augmentations, device-specific characteristics become irrelevant to the learned representation.

Sampling Rate Handling

Variable sampling rates produce variable sequence lengths for the same window duration, which must be reconciled:

1. Resampling to a fixed rate before the encoder. Simple but introduces interpolation artifacts.
2. Duration-based windowing where windows are specified in seconds rather than samples, and the model receives the actual number of samples per window as an input.
3. Implicit rate encoding where the position embedding carries the sample interval (1/SR), making the model learn that two embeddings at different rates represent the same temporal position.

Commercial Headsets

Relationship to Channel Topology Agnosticism

Device agnosticism and channel topology agnosticism are complementary: device agnosticism addresses how signals are recorded; topology agnosticism addresses where the electrodes are placed. A fully hardware-agnostic Brain FM requires both properties.