Live demo#
The interactive explainability atlas is embedded below and also lives at https://protosleepnet-demo.pages.dev. It is a static, GPU-free app: a public front-end whose derived visualizations are served from a private data store — no source recordings are shipped, only anonymized, non-invertible derived artifacts (log-power spectrograms, embeddings, prototype assignments and Integrated-Gradients maps).
Tip
If the frame does not load in your browser’s privacy mode, open it in a new tab: https://protosleepnet-demo.pages.dev.
A claim-by-claim walkthrough#
The demo’s guided Story mode walks the same path. Each panel names the claim it demonstrates (see How to interpret ProtoSleepNet for the full explanation).
The atlas is a PaCMAP projection of the 128-d embedding space with the 12 prototypes placed inside it. Because the model can only stage an epoch through its nearest prototype, this map is the model’s reasoning surface — start here.
Click any epoch. You get its true stage, its position in the night’s hypnogram, the model’s (non-quantized) prediction, its matched prototype and L2 distance, and a per-epoch Integrated-Gradients heatmap over the spectrogram — the exact time–frequency evidence for the match — with an honest plausibility badge.
Open a prototype card. Each of the 12 prototypes shows its label purity and monosemanticity score: clean, single-concept, stage-specialized patterns.
Each card’s rule, band-relevance, channel-relevance and spectral envelope follow AASM physiology — spindles for N2, delta for N3, eye movements for REM, alpha/EMG for wake.
The ribbon under each recording renders the whole night as its sequence of activated prototypes — a representation between the raw PSG and the coarse hypnogram.
No accuracy cost vs. the black-box backbone (E) and cross-dataset transfer and disease biomarkers (F) require multiple cohorts and are covered in the paper — the demo does not fabricate single-cohort visuals for them.
What is (and isn’t) shipped#
The featured recordings are the cleanest subjects from each model’s own training split, shown as anonymized “Recording A–D” with no dataset name and no demographics. Only derived, non-invertible artifacts are published; the atlas coordinates come from a stratified sample of the training embeddings, and the nearest-prototype agreement is reported honestly on-screen.