Signal and position, together.

A useful neural interface has to connect two different worlds: the electrical activity we can measure at the scalp, and the actions a person wants to take on a computer. Our work begins with the representation between them.

SetuMontageNet is a research architecture for motor-imagery EEG. It brings the structure of each recorded signal together with the location of the electrode that measured it. The aim is to make the physical measurement an explicit part of the model.

The approach is inspired by a recent line of EEG foundation models that give every electrode its own position in space. Because the network is told where each signal came from, a single model can be pretrained on very large pooled collections of recordings, gathered with many different caps and layouts, and then applied to an electrode arrangement it has never seen. We take the same idea, position as a first-class input, and study it at the small end: sparse montages, wearable form factors, and one person at a time.

Signal and position, together

Every electrode provides a view of activity from a particular place on the scalp. SetuMontageNet represents that view through temporal signal features, spectral information, and nominal electrode coordinates in a common head frame.

A shared encoder processes the measured channels. Attention combines information across the available electrode set, producing a recording-level representation for a motor-imagery decoder. The interface keeps signal values, electrode locations and channel availability aligned.

This lets us study different sets of measured channels within one network architecture. The research question is how that structure can support useful representations when the recording setup changes, and when the person wearing it does.

Learning from the available measurement

The architecture explicitly represents which channels are present. Absent channels are masked out of attention and aggregation, so the encoder builds its representation only from the electrodes that were measured.

This gives a clear basis for studying sparse montages: how electrode selection affects decoding, what information is retained as the available set changes, and how much personal adaptation a given recording setup requires.

Our focus is the relationship between the measurement and the learned representation. Keeping that relationship explicit lets us ask precise questions about what the model uses and how it responds when the input changes.

From representation to interaction

Motor imagery is our starting task. The research weights were trained within Setu Labs on the public BNCI2014-001 recordings, with outputs corresponding to that dataset’s imagery classes. The model is built from established neural-network components around an interface that includes electrode geometry.

The wider software stack separates acquisition, signal preparation, decoding and command policy. A decoder supplies class probabilities; a downstream policy assigns computer actions. This leaves room to study representation learning and interaction design independently, with explicit connections between them.

Recorded EEG provides a repeatable way to inspect the software path. Local inference and integration with a compatible wearable recording system are part of the design direction. Our longer-term goal is an interface that learns with its wearer.

The questions guiding the work

We are interested in how electrode geometry should enter an EEG model, how a compact representation changes with the available montage, and where personal calibration contributes information that pretraining alone cannot provide.

For computer interaction, these questions extend beyond classification. A useful system must also handle periods with no intended command, changing signal quality, and the feedback a person receives while using the interface. These are research objectives for the broader system.

Opening the architecture

We will open-source the SetuMontageNet architecture and its implementation. We plan to accompany the release with the model input specification and instructions for reproducing the evaluation. Release details will be announced here.

We will update this page with accuracy and other evaluation metrics, together with the dataset splits, preprocessing, baselines and experimental conditions needed to interpret them. Public-dataset evaluations and live-hardware studies will be reported separately.

This announcement introduces the architecture and the research direction. A subsequent evaluation report will present the results and their scope.

We welcome technical discussion with researchers working on neural decoding, sparse EEG acquisition and human–machine adaptation.