BCILattice is the NeuroMatrix desktop application for BCI and neuroscience research. One local workspace covers the whole study: import recordings, preprocess signals, design paradigms, build and train ML pipelines, track experiments, inspect model behavior, and export reports.

BCILattice is the desktop tool you run on your own machine, and it requires a BCINexus account for access. BCINexus also provides the connected layer around it: plan limits, cloud backup, the community study catalog, public publishing, private share links, reviewer workflows, and team workspaces.

No. The whole pipeline is visual: import data, configure preprocessing, design Neural Flow paradigms, wire up ML Suite blocks, run training in Workflow, compare runs, and export reports, all without code. Custom code blocks are there when you want lower-level control, not as a requirement.

No. There is no offline mode, and an active account is required to run BCILattice. The compute-heavy work is all local — importing, preprocessing, pipeline building, training, analysis, and report generation run on your machine — but the connection still has to be there.

Paid features come from a licence check against the server, never from cached data on disk. If the connection drops mid-session the app keeps running and paid features survive for the remaining life of the current licence token, roughly ten minutes; after that it falls back to Free features and retries in the background, restoring them on the first success. Starting with no connection at all means a Free session until you get one.

Internet is also needed for cloud sync, community publishing and imports, update checks, and remote AI providers in AI Chat.

EEG, fNIRS, EMG, and EOG each have a dedicated import and preprocessing path. EEG, fNIRS, and EMG are the most developed research flows; EOG focuses on filtering, annotation, and feature extraction.

BCI researchers, neuroscience labs, students, ML engineers, and teams that want a reproducible local workflow without rebuilding the import, preprocessing, training, and reporting stack from scratch.

A modern desktop or workstation with at least 8 GB RAM (16 GB+ for large multi-subject datasets), a multi-core CPU, and disk space for your recordings plus derived outputs. A CUDA-capable NVIDIA GPU speeds up deep-learning training but is not required, CPU training is always available.

No, for packaged builds, the installer bundles the application runtime. You only manage a Python environment if you run BCILattice from a developer source checkout, in which case use the study's environment and dependency files.

BCILattice is a PySide6 desktop application, so it is cross-platform by design. The Download page is the source of truth for which packaged installers are currently published. If your platform is not listed there, it is a developer/source setup until an installer ships.

The desktop UI talks to a local FastAPI service running on your own machine. If that service has not finished starting, API-backed pages may wait, show an empty state, or ask you to retry. Restart BCILattice first. If it keeps happening, check the app logs and make sure no other process is holding the local server port.

Day-to-day use does not require admin rights. Whether installation does depends on the package and install location. If your institution locks down software installs, use a user-level install or ask IT to approve the package.

By modality: EEG, EDF, BDF, GDF, FIF, VHDR, SET, CNT, TXT, CSV, XLSX. fNIRS, SNIRF, FIF, TXT, CSV, XLSX. EMG, EDF, BDF, GDF, FIF, TXT, CSV. EOG, EDF, BDF, FIF, TXT, CSV, XLSX.

Yes. Data Manager has folder import: point it at a directory, and it scans compatible files, preserves your subject/session structure, and lets you assign modalities and labels before adding everything to the study.

No. Preprocessing is non-destructive. Source files are never modified, derived data, preprocessing settings, study metadata, and output artifacts are all written separately.

Local studies use .blts. It stores study metadata and the workflow state BCILattice needs to reopen your study, and it is the file to move when you carry work between machines. Older session/share exports may still open, but new work should use the study workflow.

Imported recordings and derived outputs stay on your machine. Trained model artifacts live in a global per-user store (~/.bcilattice/experiments) and are referenced by experiments; the .bciproj bundle pulls them in when you export or share a study.

Yes. Reports export as PDF, HTML, JSON study documents, and CSV metric tables. Experiment Hub can also export run records as CSV for analysis outside BCILattice.

Create or open a study → import data in Data Manager → configure Preprocessing → explore in Analysis → design a paradigm in Neural Flow → build the pipeline in ML Suite and assign data in Workflow → train → review Experiment Hub and Results → inspect Model Analysis → generate a report or publish.

The entry point for files, subjects, sessions, modalities, labels, and quality checks. Everything downstream, preprocessing, analysis, workflow assignment, and training, reads from the records you define here.

The visual paradigm and stimulus designer. You lay out task timing, cues, trial markers, motor-imagery or cognitive blocks, rest periods, and loops, then compile a paradigm definition you can pair with recordings or session labels.

The visual machine-learning area. It bundles MLFlow graph editing, Workflow data assignment, training dashboards, curated BCILattice blocks, and a large catalog of scikit-learn, PyTorch, MNE, preprocessing, and evaluation components.

Experiment Hub is your run history, it stores every completed training run so you can compare, inspect, export, and restore experiment state. Results focuses on the metrics themselves and cross-experiment comparison. Use Experiment Hub to ask "what changed between runs?" and Results to ask "which run is best, and by how much?"

The interpretability surface: SHAP attributions, learned weights, and activation maps where the selected model and data support them. It is about understanding why a model behaves the way it does, not just its score.

Yes, the workspace includes Record Session (live streaming acquisition) and Edge Devices (on-device model deployment) surfaces. Both are marked experimental in the app and not yet validated against physical acquisition hardware, so verify channel labels and sample rates against your amplifier before relying on a recording. They are optional local tools; analysing imported datasets does not need them.

The catalog includes BCI-specific EEG architectures, EEGNet, ShallowConvNet, DeepConvNet, EEGConformer, and ATCNet, alongside sequence and transformer models, edge-focused models, classical estimators (LDA, SVM, random forests, gradient boosting, and more), feature selection, dimensionality reduction, evaluation metrics, and custom blocks.

Training runs on your machine. PyTorch models use a compatible NVIDIA GPU when one is available and fall back to CPU otherwise. The actual speedup depends on your installed runtime, drivers, and the chosen model. There is no cloud GPU, BCINexus does not host training.

Keep subject/session splits explicit in Workflow and use validation or cross-validation blocks that match your study design. Never fit scalers, feature selectors, or dimensionality reduction on the full dataset before splitting, fit them inside the training fold only. For subject-generalization claims, use subject-aware validation such as leave-one-subject-out.

Yes, on the Lab plan and up. Export runs from the Model Analysis “Export Model” action or the desktop edge workflow for supported experiments. PyTorch and ONNX are the main portable targets (scikit-learn and other native formats are supported too), with edge-specific options where the required conversion tools are installed. Free and Researcher plans do not include model export or edge deployment.

Yes, there are custom-code entry points for preprocessing, feature extraction, and pipeline work. Document your dependencies, seeds, and data splits when you do: undocumented custom code is the easiest way to make a study hard to reproduce.

AI Chat connects to your own provider accounts: Gemini (Google), ChatGPT (OpenAI), Claude (Anthropic), Mistral, DeepSeek, a free-tier Groq API, and a local Ollama endpoint. You supply the API key (or run Ollama locally); BCILattice does not resell access.

Straight from your desktop to the provider endpoint you selected, using your own credentials. The BCINexus backend is never in the loop. Treat any prompt to a remote provider as data leaving your machine, use the local Ollama endpoint when prompts must stay on-device.

It is grounded in your work through retrieval over app knowledge, your experiments, sessions, and pipelines, so you can ask it to explain a preprocessing choice, interpret a result, or troubleshoot a pipeline with context from your own study, not just generic answers.

Yes. A BCINexus account is required to operate BCILattice. Local analysis and training still run on your machine; cloud sync, private sharing, team workspaces, community publishing, reviewer workflows, and dashboard features are connected account services.

Yes. Browse the community catalog on BCINexus or from the desktop Community entry points, then import it into a study, as a new study or into the current one. Always review the imported graph and settings before running it on your own data.

Use the desktop Publish flow from the study workspace or the Import / Export page. You choose the title, description, modality, task type, visibility, and which artifacts to include. Published items are versioned and tracked through the BCINexus dashboard.

Yes. Private cloud save and share-link flows are available to signed-in users. A private link grants access to the selected study or package without listing it publicly in the community catalog.

A desktop-side view of your BCINexus publish/review state, so you can see whether a submitted item is pending, accepted, rejected, or needs follow-up without leaving the app.

These are the live backend quotas:

Community downloads are unlimited on every plan, including Free — we do not charge anyone to read what someone else chose to publish. Upload allowances reset on the 1st of each month.

An "upload" is a published or cloud-stored item; downloads from the community are metered weekly on the Free plan only.

Team plans (Lab and above) unlock shared workspaces: members and roles, shared studies and files, dataset manifests, tasks, pages, activity logs, and study requests, all drawing on the team's quota instead of an individual's.

No. Training always runs on each researcher's own machine. BCINexus shares study artifacts, files, dataset manifests, and collaboration state, it is not a hosted GPU service.

Yes, route Enterprise arrangements through the sales/enterprise contact flow, since legal review, deployment model, support terms, and billing vary by organization.

No. Imported recordings stay local. Data leaves your machine only when you explicitly run a cloud sync, publish, share, or AI-provider action that includes it. This separation matters for IRB, clinical, and restricted-research settings.

Whatever you put in the prompt, sent directly to the provider you connected, using your credentials. Assume prompt content leaves your machine unless you are using the local Ollama endpoint.

It supports local-first, institution-controlled workflows, but compliance depends on your deployment, policies, agreements, and data handling. The product does not carry a blanket HIPAA, GDPR, SOC 2, or clinical certification, review the Security & Compliance page and involve your compliance team.

Use account settings for self-service controls. For privacy or right-to-erasure requests, email [email protected].

Confirm the extension matches the real file type, open it in a trusted external tool, and make sure the selected modality is correct. For CSV/TXT/XLSX, check that the table is numeric and the label column is clearly identified.

Start small: fewer folds, a lighter model, and a sanity check on your data shape. Confirm the GPU runtime is actually being used. When debugging a new dataset, run a classical baseline (LDA or SVM) before reaching for deep models.

Almost always an incomplete graph, a missing required parameter, an incompatible connection, or an output block that doesn't match what training expects. Read the graph validation message before changing the model.

Sign-in state, active workspace, plan quota, internet connection, and whether the selected artifacts are valid. If a team workspace is active, quota and permissions come from the team, not your personal plan. An expired session needs a fresh sign-in, refreshing only works before expiry.

Email [email protected] with your app version, operating system, the module where the issue happened, and the relevant error text from the desktop logs.