A selective dual-path sensor encoder maintained decision accuracy (AUC 0.674) compared to a serial transform (AUC 0.610) while allowing exact waveform reconstruction for selected windows.
A novel dual-path sensor encoder allows extreme data compression while preserving decision accuracy and enabling on-demand lossless waveform reconstruction for selected windows.
Absolute Event Rate: 0.674% vs 0.61%
A class-discriminant codebook compresses each window of a sensor signal to a single decision-oriented token; the token preserves the decision but is, by construction, not invertible to the waveform. Yet some windows — a flagged arrhythmia for clinician adjudication, a developing fault for forensics, an audited replay — later require the exact waveform. Transmitting the losslessly-coded waveform for every window defeats extreme compression; front-loading a re-coordinatizing transform before the codebook degrades the decision. We disclose a selective, tiered-fidelity, dual-path encoder: for each window an always-on decision path (the codebook token) and an on-demand lossless reconstruction co-channel (a bijective logarithmic spiral-domain transform) operate in parallel on the same raw window, and a router invokes the co-channel only for windows selected by a monitor derived from the decision path itself (the codebook's nearest-centroid anomaly score). On real public benchmark data, with every outcome band frozen before analysis, we show: the parallel arrangement leaves the decision unchanged (AUC 0.674) whereas a serial front-loaded transform degrades it (0.610); the spiral path reconstructs the raw waveform to machine precision (mean absolute error of order 1e-9 on ECG, 1e-18 on industrial vibration); routing by the codebook's own monitor preferentially selects diagnostically-relevant windows (selection AUC 0.69–0.70 versus 0.50 random), and an ensemble with a reconstruction-residual gate improves selection over either component alone; the architecture is cross-modality (industrial selection AUC 0.946); and the exact reconstruction supports operations the token cannot — a continuous waveform measurement (token-to-reconstruction error ratio of order 1e11) and fine-grained sub-typing of a flagged abnormality (0.783 vs 0.645 from the token). The economy is on the transmission and storage axes (lossless payload reduced in proportion to the routing budget; ≈20×/10×/5× at 5/10/20% routed), not computation (the dual-path costs ≈1.02× the decision path). We report three honest negatives verbatim: a finely-graded multi-tier ladder does not materialize on high-rank raw signals (a coarse intermediate tier only, for both a linear and a residual-vector-quantization code); the co-channel does not materially improve prediction (a remaining-useful-life correlation moves only 0.749→0.754); and an anomaly-distance-only router systematically under-routes near-manifold classes (ST/T-change windows at recall 0.07), which class-aware and ensemble routing mitigate. The subject matter is filed (Parent T, 64/098,837); the contribution is the selective dual-path that delivers the decision for every window and the exact waveform only where a monitor says it matters. Keywords / index terms: tiered-fidelity coding; selective reconstruction; dual-path encoder; class-discriminant codebook; logarithmic spiral-domain transform; lossless reconstruction; anomaly-routed selection; on-demand reconstruction; edge sensing; bandwidth economy; class-aware routing; pre-registration; honest negatives. References (selected): 1. Y. Linde, A. Buzo, R. M. Gray, "An algorithm for vector quantizer design," IEEE Trans. Communications, 1980. 2. N. Tishby, F. C. Pereira, W. Bialek, "The information bottleneck method," Allerton, 1999. 3. H. Schwarz, D. Marpe, T. Wiegand, "Overview of the scalable video coding extension of the H.264/AVC standard," IEEE TCSVT, 2007. 4. D. Gündüz et al., "Beyond transmitting bits: Context, semantics, and task-oriented communications," IEEE JSAC, 2023. 5. Y. Geifman, R. El-Yaniv, "Selective classification for deep neural networks," NeurIPS, 2017. 6. D. Hendrycks, K. Gimpel, "A baseline for detecting misclassified and out-of-distribution examples," ICLR, 2017. 7. R. J. Ferlic and K. K. Ferlic, companion deposits (Papers 19–26), Zenodo, 2026. Companion deposits in this Zenodo Community (spiral-domain-encoder-campaign): · Paper 19 — 10.5281/zenodo.20788187 · Paper 20 — 10.5281/zenodo.20802759 · Paper 21 — 10.5281/zenodo.20802826 · Paper 22 — 10.5281/zenodo.20805321 · Paper 23 — 10.5281/zenodo.20821668 · Paper 24 — 10.5281/zenodo.20821779 · Paper 25 — 10.5281/zenodo.20821903 · Paper 26 — 10.5281/zenodo.20854722
Ferlic et al. (Thu,) reported a other. Selective tiered-fidelity dual-path sensor encoder vs. Serial front-loaded transform was evaluated on Decision AUC. A selective dual-path sensor encoder maintained decision accuracy (AUC 0.674) compared to a serial transform (AUC 0.610) while allowing exact waveform reconstruction for selected windows.