BVP, EDA and temperature traces stream in from the wrist.
Research Overview
Key Results
From Raw Biosignals To An Affective Read In Three Steps
Each physiological stream is treated as a sequence of irreducible atoms. The engine checks that every transition makes sense, repairs the ones that don’t, and scales its own reasoning to the difficulty of the interval.
-
01
Atomize
Per-sensor TCNs, cross-modal attention, and a Variational Information Bottleneck transform BVP, EDA, and temperature into compact physiological atoms that preserve affect-relevant autonomic dynamics while suppressing noise.
-
02
Bridge
The Markovian Reasoning Engine detects logic gaps and inconsistent multimodal transitions. Context-conditioned diffusion synthesizes bridge atoms to restore continuity, while high-integrity contradictions are retained as novel patterns for classifier fine-tuning.
-
03
Scale
At inference, Markovian test-time scaling adapts diffusion depth to contradiction severity, enabling targeted self-correction on difficult or noisy intervals without the memory overhead of conventional temporal models.
Interactive Reasoning Console
An interactive visualization of AtomSigNet’s atom-level reasoning, contradiction detection, diffusion-based repair, and affect prediction across time.
Contradictory Emotions Across Modalities
The console reads the visual, acoustic and physiological streams separately and shows where they conflict instead of collapsing them into a single label.
AtomSigNet: Reading Biosignal One Atom At A Time
The wearable stream is broken into short atomic units. A diffusion model fills in atoms that don’t fit the sequence around them, and at test time the model keeps refining until the sequence holds together before the affect classifier makes its call.
Real-Time Multimodal Inference Stack
Implementation details of the end-to-end sensing, affect inference, and visualization pipeline. Core video, audio, and physiological inference runs locally. Only the optional natural-language summary is generated externally.
State Of The Art On WESAD And CASE
AtomSigNet outperforms existing supervised and self-supervised models on both benchmarks, and fine-tuning with synthesized samples amounting to less than 1% of the training set is already competitive with state-of-the-art baselines.
Acknowledgements
Citing This Work
If you use or build upon this work, please cite the corresponding publication(s) from our related work listed below.
References
Datasets, methods and models referenced on this page. If you build on these resources, please also cite the original works.