BCI Oscillatory Encoder

Export
LOCAL BROWSER ENGINE
ReadyNo run loadedSimulated oscillatory templates · no measured neural activity

02 / Signal analysis AWAITING RUN

Visual input 64 × 64 RGB

Run the encoder to inspect input frames
Synthesized waveform · all bandsAmplitude (a.u.)
Window time (ms)No samples
Template band powerPower coefficient (a.u.)
Input time (s)No samples
Wrapped phasePhase (rad)
Input time (s)No samples
Spatial FFT energyEnergy fraction
Input time (s)No samples
Temporal spectrum of frame motionNormalized power (a.u.)
Temporal frequency (Hz)No samples
Visual feature trajectoryNormalized feature
Input time (s)No samples

Verification

Python reference fixtures

The regression suite compares browser computations with NumPy, OpenCV, and Torch on shared RGB fixtures. It checks visual features, temporal spectra, band coefficients, phase, spatial weights, waveform samples, RMS and peak.

Feature absolute tolerance
2 × 10⁻⁶
Template absolute tolerance
1 × 10⁻⁶
Waveform absolute tolerance
5 × 10⁻⁶

These checks assess numerical agreement on fixed inputs. Decoding and resizing can vary by browser. The experiment does not validate a neural response.

Loaded run provenance

No run loaded.

Methods

This experiment asks how visual sequences can be represented as compact, multiband oscillatory templates for computational analysis.

Visual features

The browser decodes and resizes each frame to 64 × 64 before grayscale conversion. Codec, color conversion, and resampling can differ from the Python OpenCV pipeline. Video analysis uses uniformly spaced seek times at the configured rate. The analysis rate describes those sample times. Eight measurements describe luminance, contrast, Canny edge density, low and high non-DC spatial FFT energy, center and peripheral luminance, and mean absolute frame difference. Three mean-centered, L1-normalized cosine–Gaussian kernels at scales 1, 2 and 3 produce eight temporal features by averaging the absolute filter responses. The workspace also displays the squared normalized FFT magnitude of the centered motion trajectory, using the input frame rate as the timebase.

Deterministic mapping

The 6, 10, 20 and 40 Hz bands use motion, brightness, contrast and the mean of edge density and high-frequency energy as respective power drivers. Each driver is scaled by 0.6, offset by 0.15 and multiplied by a channel gain from 0.8 to 1.2. Phase maps bounded motion, brightness, contrast and edge density linearly to [−π, π]. Spatial weights normalize a center–periphery mixture across channels.

GRU and circular phase

A 16-feature input projection with ReLU feeds a 64-unit GRU. Three heads emit nonnegative band power through softplus, normalized phase pairs converted to angles by atan2, and softmax channel weights. Phase error uses the mean of 1 − cos(Δφ) across the circular wrap. The held-out synthetic benchmark compares power MSE, circular phase error and spatial MSE against a constant predictor. This benchmark and checkpoint-based comparisons are available in the local Python implementation.

Synthesis and interpretation

For each input frame and channel, synthesis sums Pᵦ sin(2πfᵦt + φᵦ) over the fixed bands and multiplies by the channel weight. Template power is used as the amplitude coefficient by convention. Each frame creates an independent window. Phase is circular within the representation, and waveform continuity across windows is unconstrained.

Evaluation measures reproduction of a synthetic deterministic teacher. Visual-to-oscillatory representations have not been validated as a means of delivering visual perception to the human brain. The workbench is computational only and produces simulated oscillations in arbitrary units.

Research and collaboration

This experimental framework is available for research exploration. The project welcomes technical evaluation, research collaboration and inquiries about potential nonexclusive software licensing.