Autara-OF: High-Performance GPU-Accelerated BCI Classifier

Autara-OF is a highly generalized, hardware-accelerated Brain-Computer Interface (BCI) neural network. It utilizes an early-fusion Multi-Modal architecture to decode human intent by mathematically bridging the rapid electrical firing of neurons (EEG) with deep, localized metabolic blood-oxygen flow (fNIRS).

Architecture Details

The model relies on a deeply correlated Transformer Cross-Attention block to merge the two independent biological modalities:

• EEG Encoder: 8-Channel Conv1D Network mapping high-frequency electrical signatures.
• fNIRS Encoder: 16-Channel Conv1D Network mapping slow-wave hemodynamic oxygenation.
• Fusion Layer: Cross-Attention matrices projecting EEG query spaces into fNIRS key/value pairs to extract deep contextual human intent.

Dataset & Training Constraints

• Data Source: Trained against a massively augmented 10GB subset of OpenNeuro's ds007554 clinical trial.
• Resolution: 60,481 deep arrays (200 timesteps spanning 5-seconds of human thought).
• Optimization: Converged using AdamW bound by severe weight-decay (0.01) and a Cosine Annealing Learning Rate trajectory to prevent outlier gradient explosions.

Clinical Real-Time Capabilities

• Task Classification: Distinguishes between Active Motor (physical/imagined movement) and Mental Arithmetic (complex internal cognition).
• Latency: Sustains <1.0 ms inference speeds natively on an NVIDIA RTX 3070.
• Accuracy: Locks into unseen human biological vectors with 99.99% Softmax Confidence in strictly isolated testing loops.

Usage

The autara_of_weights.mpk binary is compiled exclusively for the burn-rs Deep Learning framework.

// Restore Graph
let record = NamedMpkFileRecorder::<FullPrecisionSettings>::new()
    .load("autara_of_weights".into(), &device)
    .expect("Failed to decode weights");

let model: AutaraOFModel<B> = config.init(&device).load_record(record);