tqnnlabs/Tqnnlabs-Benchmarks

TQNN Labs Benchmarks

This repository contains public benchmark artifacts for TQNN Labs experiments, including subject-level EEG/BCI robustness comparisons.

TQNN v5.8 EEG/FBCSP Robustness Benchmark

This benchmark compares TQNN-Q feature inference against standard FBCSP-based EEG pipelines across 9 subjects and 27 subject-seed runs.

FBCSP stands for Filter Bank Common Spatial Pattern, a commonly used EEG/BCI feature extraction approach.

Summary Result

Pipeline	Mean Accuracy
Full FBCSP LR	51.23% ± 7.94%
Full FBCSP LDA	50.00% ± 7.89%
Selected-13 FBCSP LR	51.13% ± 8.01%
Selected-13 FBCSP LDA	51.23% ± 7.91%
TQNN-Q LR	49.69% ± 8.92%
TQNN-Q LDA	50.31% ± 8.40%

Main Comparison

TQNN-Q LDA vs Full FBCSP LDA

• Mean improvement: +0.31% ± 8.99%
• Win / tie / loss: 16 / 3 / 8
• Total completed subject-seed runs: 27
• Subjects completed: 9 / 9

Interpretation

The result is intentionally modest and robustness-focused.

TQNN-Q LDA performed competitively against a full FBCSP LDA EEG pipeline across all tested subject-seed runs, with a small positive mean improvement and more wins than losses.

This does not claim quantum advantage.

It shows that the TQNN selected-13 quantum-classical feature path can remain competitive with a dedicated EEG signal-processing baseline while using a compact structured representation.

Why this matters

EEG is a noisy, difficult signal domain. A general-purpose architecture holding up against a specialized EEG pipeline is an important systems result.

TQNN is designed as a structured-data inference layer, meaning the same architecture can be applied across:

• EEG / BCI
• Finance
• Chemistry
• Tabular data
• Text-derived features
• Image-derived features

The goal is not to replace every domain pipeline immediately. The goal is to test whether a single quantum-classical inference substrate can produce useful, stable representations across many structured data types.