About This Reference

This document covers every component available in BCILattice's three designer environments: NeuralFlow (timeline/paradigm designer), MLFlow (ML Suite node-graph), and the Workflow Designer (orchestration graph).

Part 1, NeuralFlow Block Types

NeuralFlow is BCILattice's timeline-based paradigm designer. Blocks are arranged left-to-right to define the temporal sequence of a BCI experiment. Each block represents one epoch type, a control marker, or a structural element.

Every block has three editable properties: Label (display name), Label Code (integer sent to EEG hardware as epoch marker), and Duration in seconds (1–600). Special control blocks have reserved codes.

All 28 Block Types

Loop Block, Detail

The Loop Block is a container that repeats a set of sub-blocks. It cannot be nested inside another Loop Block. Duration is auto-computed as: sum(sub_duration × sub_repeat) × total_repeat.

Part 2, MLFlow (ML Suite) Components

The ML Suite canvas is a node-graph pipeline builder. Drag blocks onto the canvas, connect ports, then Compile and Train. All blocks have typed input/output ports that are validated at compile time. The catalog is loaded from pipeline_catalog_final2.csv at runtime.

Pipeline Blocks (Built-in)

These blocks are injected at runtime, not from the CSV catalog. Available in both MLFlow and Workflow.

Analysis Blocks (Custom)

BCI EEG Models

Custom EEG architectures built with PyTorch. Module: nm_models.bci_eeg.

Sequence & Transformer Models

Edge & GPT Models

Feature Selection, sklearn.feature_selection (19 blocks)

Cross Validation, sklearn.model_selection (32 blocks)

Dimensionality Reduction, sklearn.decomposition (33 blocks)

ML Linear, sklearn.linear_model (44 blocks)

Classifiers and regressors based on linear models. Key BCI-relevant blocks:

ML Ensemble, sklearn.ensemble (48 blocks)

ML Neighbors, sklearn.neighbors (24 blocks)

ML Clustering, sklearn.cluster (29 blocks)

Preprocessing, sklearn.preprocessing / sklearn.impute (32 blocks)

Evaluation, sklearn.metrics (84 blocks)

Key metrics for BCI classification evaluation:

Model Interpretation, SHAP (63 blocks)

Module: shap. All 63 blocks provide SHAP-based explainability, attributing feature importance to individual predictions. Key explainers:

• TreeExplainer, for tree-based models (Random Forest, Gradient Boosting)
• DeepExplainer, for PyTorch/TensorFlow deep learning models
• LinearExplainer, for linear models (LDA, Logistic Regression)
• KernelExplainer, model-agnostic, any estimator
• GradientExplainer, gradient-based explanations for PyTorch models
• Explainer, unified auto-selecting explainer
• Visualisations: summary_plot, waterfall_plot, force_plot, bar_plot, beeswarm_plot

Signal Processing (MNE), 273 blocks

Module: mne. Key block groups:

Signal Processing (General), 228 blocks

Libraries: pywt, scipy.signal, librosa.

• PyWavelets, dwt, wavedec, wavedec2, ContinuousWavelet, DiscreteWavelet, tree nodes
• SciPy signal, spectrogram, stft, hilbert, butter, sosfilt, filtfilt, correlate
• librosa, mfcc, chroma_stft, mel_spectrogram, spectral_centroid, zero_crossing_rate

Large Catalog Categories

Part 3, Workflow Designer Components

The Workflow Designer shares all MLFlow blocks and adds three workflow-only blocks for file selection, REST API exposure, and result saving. It orchestrates multi-step processes across entire datasets rather than single training runs.

Workflow-Only Custom Blocks

A typical complete workflow graph follows this pattern:

1. Select Files → choose sessions, modality, and data source
2. Channel Selection → rank and remove low-discriminability channels
3. Preprocessing blocks → StandardScaler, feature extraction
4. ML Pipeline → run a compiled MLFlow pipeline per file
5. Evaluation blocks → accuracy, confusion matrix, F1 scores
6. Graph Output → save result plots to disk
7. Graph Endpoint (optional) → expose results via REST API

Summary