SETU Neural EEG
Trained models for the next human interface.
Today, we’re releasing SETU Neural EEG v0.1.0: our trained EEG model suite and the software behind it.
The release brings together task-trained neural networks, fine-tuned EEG encoders and command classifiers, supported by a common training, adaptation and evaluation pipeline.
Our ambition is to build the intelligence layer that makes neural interaction useful: models that learn from brain signals, adapt to people and connect human intent with digital systems.
Intelligence is advancing. The interface should advance with it.
AI is changing what computers can do. We believe neural interfaces can change how people access that capability.
Today, most digital interaction passes through a narrow set of inputs: keyboards, touchscreens, physical controllers and speech. Each places demands on the person using it.
Our vision is to expand that interface.
Imagine navigating software through an intended command. Accessing communication tools when conventional inputs are difficult. Interacting with an intelligent system through a new channel of expression.
We want neural interaction to become a practical part of computing.
Setu Labs begins with the models: learning useful representations from EEG, adapting them to a task and turning recorded signals into command predictions.
Our models. Our training. A public release.
SETU Neural EEG is the result of a broad downstream training and adaptation effort across neural and classical approaches.
We trained compact EEGNet and EEG Conformer networks for imagined-command classification. We trained task-specific heads and fine-tuned pretrained EEG encoders. We fitted support-vector machines, covariance-based classifiers and filter-bank pipelines.
We then evaluated these approaches within a shared framework, across individual participants and different electrode configurations.
This work produced a suite of trained artifacts that others can download, inspect and run.
The release puts the models themselves in your hands.
Built around the whole decoding pipeline.
A neural interface depends on more than a model architecture.
The signal has to be prepared correctly. The model has to adapt to the person. Sensor choices have to be reflected in training. Predictions have to be evaluated under a clear protocol.
We built the Setu Labs pipeline around those requirements:
- Signal processing: trial reconstruction, channel selection, rereferencing, filtering and resampling.
- Model training: task-specific neural networks, fitted classifiers and pretrained-encoder adaptation.
- Participant calibration: trained models and selected hyperparameters for individual benchmark participants.
- Channel studies: retrained configurations using 8, 16, 32 and 64 measured electrodes.
- Inference and evaluation: reusable loaders, command prediction, frozen model selection and complete comparison records.
Together, these components form the technical foundation for our next stage of neural-interface development.
A concrete result. A platform for the next advance.
On the strict five-command imagined-speech benchmark, our validation-selected 32-channel decoder achieved 30.4% held-out accuracy, against a 20% uniform-chance baseline. The evaluation covered 750 held-out trials across 15 calibrated participants.
That result gives us a measurable starting point. The broader model suite gives us a way to investigate what improves it: different representations, adaptation methods, sensor configurations and calibration strategies.
We are publishing the complete results alongside the release so that each subsequent advance can be assessed against a visible baseline.
Open the work. Expand what people can build.
We’re releasing SETU Neural EEG for researchers, engineers and collaborators working on the next generation of neural interfaces.
The release includes:
Trained checkpoints
Eligible neural models, adapted encoders, fitted classifiers, probes and experimental controls.
Training and inference code
The software used to prepare signals, train models and generate predictions.
Experimental records
Configurations, training histories, participant-level results and model-selection records.
Reproducibility artifacts
Checkpoint hashes, loading examples, evaluation audits and documented provenance.
We verified the integrity of every released checkpoint, tested inference across representative configurations and checked that the selected model can be downloaded and run from the public release.
Our goal is to give the field something concrete to build on.
The roadmap is a more natural interface.
Our next phase targets the capabilities that matter for practical neural interaction.
Less calibration
Reduce the effort required to adapt a system to a person.
Cross-session reliability
Improve consistency across days and changing recording conditions.
Simpler sensors
Carry model development toward practical electrode configurations and prospective wearable testing.
Intentional interaction
Progress toward systems that recognize both intended commands and the absence of a command.
These priorities connect our model work to a larger product vision: neural interfaces that fit into people’s lives.
Building toward a new layer of computing.
We see neural interaction as a meeting point for advances in AI, sensing and human-centered design.
Our long-term ambition is to build software that can adapt across people and devices, and help turn improvements in neural decoding into useful experiences.
That ambition creates a shared challenge for scientists, developers, hardware builders and partners. It also creates an opportunity to expand access to computing in ways conventional interfaces cannot.
SETU Neural EEG is our first public model release on that path.
The models are available. The code is open. The work continues.
Setu Labs: intelligence for the next human interface.
Research release. Current results concern retrospective, cued five-command classification in previously calibrated participants; they do not establish open-vocabulary thought decoding, live wearable performance, clinical utility or calibration-free generalization. Cue-related information and non-neural artifacts remain possible contributors. Accuracy: 30.4%; participant-bootstrap 95% CI: 27.47–33.47%. Checkpoint counts include participant-specific fits, probes and controls, not distinct architectures. Setu Labs trained downstream models and fine-tuned pretrained encoders; upstream foundation architectures were not pretrained by Setu Labs. Code is open source; weights retain their applicable licenses. CSBrain weights and Chisco trained artifacts are withheld pending redistribution clarification. Results · Licenses and provenance.