PaperlayerView exact revision on arXiv

Source-linked AI summary

WiMoT: Replay-Free Continual Wi-Fi Sensing via Moment Transport

Chunchun Yang, Yang Cao, Zhitang Chen

arXiv:2610.32660v2cs.AI

TL;DR

Continual Wi-Fi sensing must learn across changing rooms, days, and users while retaining earlier recognition, even when privacy and storage constraints limit keeping CSI samples. WiMoT stores and transports class moments while combining supervised contrastive learning, a historical loss, and moment-based classification. It forgets at most 0.58 pp across three streams and leads the strongest baseline on 3DO and Widar3.0.

  • Problem

    Continual Wi-Fi sensing must adapt to changing rooms, days, or users while retaining earlier recognition, but privacy restrictions and limited edge storage can make CSI retention difficult.

    Source: §I · p. 1

  • Method

    WiMoT stores class means and covariances, transports them through affine updates on a frozen basis, and uses SupCon, a closed-form historical loss, and an OSCL classifier.

    Source: §I · p. 2, D. Balancing Plasticity and Stability:… · p. 3, E. A Classifier That Follows the… · p. 4

  • Results

    WiMoT forgets at most 0.58 pp across three CSI streams and has the highest average final accuracy on 3DO and Widar3.0, 3.77 and 13.11 pp above the strongest of 21 baselines.

    Source: §VII · p. 12

  • Takeaways & Limitations

    WiMoT’s stored moments keep memory valid without retaining old samples, while its classifier uses all observed domains without domain identity.

    Source: §I · p. 2

    WiMoT is less suited to short streams with little baseline forgetting, few labeled records per class relative to the basis dimension, or frozen encoders that separate classes poorly.

    Source: C. Limitations · pp. 11–12

Abstract

from arXiv · show

Continual Wi-Fi sensing must balance plasticity, learning new rooms or users, against stability, keeping earlier ones, yet nodes often cannot retain channel state information (CSI) for privacy and storage reasons. We propose WiMoT (Wi-Fi Moment Transport), which stores the per-class feature mean and covariance of every earlier domain instead of samples. When the representation changes, stored statistics become stale. WiMoT restricts the change to an affine map on a frozen feature basis and applies the same map to the statistics, so they stay valid without old samples. Supervised contrastive learning learns the new domain, a closed-form loss on the stored statistics keeps earlier classes separable, and a Gaussian discriminant classifier, refitted from the transported statistics with equal weight per domain, predicts without domain identity. On 3DO, WiMANS, and Widar3.0, WiMoT's average final accuracies are 94.02%, 89.61%, and 95.16% with at most 0.58 percentage points (pp) of forgetting. On 3DO and Widar3.0, it also has the highest balance of new- and earlier-domain accuracy and is 3.77 and 13.11 pp ahead of the strongest of 21 baselines, which uses replay, staying ahead across domain orders and label budgets. On WiMANS, it matches the strongest replay baseline in accuracy (+0.41 pp) while reducing forgetting from 2.78 to 0.00 pp, and has the highest six-order average accuracy. Because full covariances dominate the memory, storing them as rank-32 factors shrinks the moment memory to 6.4%, losing at most 0.49 pp of accuracy, within edge-node storage budgets.

I. INTRODUCTION

Continual Wi-Fi sensing must adapt across changing acquisition conditions while retaining earlier recognition, but privacy and storage constraints can rule out retaining CSI recordings. WiMoT addresses the resulting stability, drift, and classification challenges with stored class moments and a constrained representation update.

I. INTRODUCTION

  • Privacy restrictions and limited edge storage make retaining CSI recordings difficult as sensing nodes learn new rooms, days, or users.The task is domain-incremental learning with a fixed activity vocabulary.Source: §I · p. 1

B. Continual-Learning Mechanisms

  • Frozen representations preserve stored statistics but learn new domains poorly, while unconstrained updates forget and nonlinear drift estimates are approximate.On a nine-user stream, fine-tuning forgets 28.78 pp, SI 20.17 pp, and ConSense 27.50 pp.Source: §I · p. 1
  • WiMoT combines SupCon plasticity with a closed-form historical loss that uses stored moments to penalize reduced readability of earlier labels.The historical loss is the earlier domains’ squared classification loss, evaluated in closed form from moments,.Source: §I · p. 1, B. Continual-Learning Mechanisms · p. 2

C. Drift Compensation of Stored Statistics

  • WiMoT restricts representation changes to an affine adapter on a frozen basis, making the same map transport stored moments exactly without old samples.This differs from methods that estimate a drift map from data, so their transported statistics are approximate.Source: §I · p. 1, C. Drift Compensation of Stored… · p. 2

D. Statistical Continual Classification

  • OSCL builds a classifier from class moments and optimized combining labels, with its no-forgetting property conditional on a fixed feature space and valid moments.WiMoT supplies these conditions through its frozen basis and moment transport.Source: §I · p. 1, D. Statistical Continual Classification · p. 2

III. PROBLEM SETTING

  • Training uses current-domain data and stored aggregate statistics, while inference receives the input without its true domain identity.Domain boundaries and current labels are known during training; earlier individual recordings are not revisited.Source: §III · p. 2

A. Design Rationale

WiMoT uses stored class moments to address three linked needs: learning new domains, keeping earlier classes readable, and maintaining valid memory as its representation changes. It builds this system around a frozen feature basis, affine adaptation, and a historical loss.

A. Design Rationale

  • WiMoT replaces samples with class means and covariances, transporting the memory alongside an affine update and refitting a classifier on all domains with equal weight and no domain identity.Without transport, AFA falls by 12.1, 3.4, and 31.0 pp relative to transport.Source: A. Design Rationale · pp. 2–3

B. Feature Basis and Model

  • A frozen semantic and amplitude basis feeds an affine adapter; the contrastive projector and historical-loss map are training-only, while the statistical classifier operates on the adapter output.The basis combines frozen stems with a first-domain PCA shortcut.Source: B. Feature Basis and Model · p. 3

C. Keeping the Memory Valid: Moment Transport

  • For each domain and class, WiMoT stores counts, feature means, and sample covariances; the affine adapter transports means and covariances exactly without old samples or drift estimates.The identity applies to the stored statistics, not to whether earlier classes remain separable; nonlinear adaptation is limited after the first domain.Source: C. Keeping the Memory Valid: Moment… · p. 3

D. Balancing Plasticity and Stability: SupCon and the Historical Loss

  • SupCon fits the current domain, while a closed-form historical loss uses transported moments to penalize reduced linear readability of earlier labels.Linear readability is a proxy for OSCL accuracy, not a guarantee; both moment identities rely on a fixed basis and affine adapter.Source: D. Balancing Plasticity and Stability:… · pp. 3–4

E. A Classifier That Follows the Representation: OSCL

WiMoT refits OSCL from transported moments after each update so the classifier follows the current representation and includes all observed domains. Domain-conditioned OSCL posteriors are combined without supplying domain identity.

E. A Classifier That Follows the Representation: OSCL

  • WiMoT rebuilds OSCL after each update from transported moments of all domains with equal weight, rather than relying on a classifier fitted only to current samples.OSCL estimates class errors from moment-based score means and variances and optimizes combining labels; its earlier-task guarantee is asymptotic and motivates, rather than bounds, forgetting here.Source: E. A Classifier That Follows the… · pp. 4–5
  • The illustration separates plasticity from representation stability and memory stability: SupCon, the historical loss, and moment transport address these respective roles.It depicts synthetic two-dimensional points, with rings marking samples misread by the best linear map W.Source: Figure 2 · p. 5
  • A shared Gaussian model supplies domain probabilities, and summing each domain’s OSCL class posterior weighted by that probability removes the need for domain identity.A query is decided mainly by the OSCL rules of domains it resembles.Source: E. A Classifier That Follows the… · p. 5

F. Update Procedure and Complexity

Each new domain is processed by encoding arrivals, computing their class moments, training with current-domain and historical losses, transporting stored moments, and refitting OSCL. The procedure then adds the new domain’s historical constraints and deletes its samples.

F. Update Procedure and Complexity

  • Each update encodes current arrivals once, computes their class moments, trains with 0.5LSC + Lhist, transports stored moments, refits OSCL, then adds the new moments and deletes arrivals.Lhist uses the stored Bs and Ks of earlier domains.Source: F. Update Procedure and Complexity · p. 5
  • With basis width d = 512 and adapter width k = 128, evaluating Lhist costs O((t −1)C(d + 1)2) per step, independent of historical sample count.Classifier refitting costs O(tC(kd2 +k2d)) once per update.Source: F. Update Procedure and Complexity · p. 5
  • Covariance storage grows as O(TCd2), rather than with the number of samples.The update procedure retains aggregate moments, not historical individual recordings.Source: F. Update Procedure and Complexity · p. 5

V. EXPERIMENTS

The experiments compare WiMoT with replay and non-replay baselines across three Wi-Fi sensing streams. WiMoT achieves the highest average final accuracy on all three without storing samples, with low forgetting and advantages over the strongest replay baselines on two streams.

A. Setup

  • The evaluation covers 3DO acquisition days, WiMANS rooms, and Widar3.0 users, with test accuracy measured on all observed domains after each update without domain identity.The study compares 21 baseline configurations, including replay and non-replay methods.Source: A. Setup · pp. 5–6, §V · p. 5

B. RQ1: Overall Effect Without Sample Replay

  • WiMoT attains the highest AFA on all three streams without storing samples, beating the strongest replay baseline by 3.77 ± 0.67 pp on 3DO and 13.11 ± 1.22 pp on Widar3.0.On WiMANS, it matches MAS+ER (+0.41 ± 0.64 pp) while reducing AFR from 2.78 to 0.00 pp.Source: B. RQ1: Overall Effect Without Sample… · p. 6, Table I · p. 6
  • WiMoT’s AFR is 0.32, 0.00, and 0.58 pp on 3DO, WiMANS, and Widar3.0, respectively, versus 5.64, 2.78, and 10.25 pp for the strongest baselines.Its advantage is largest on the longest stream, Widar3.0, where baseline forgetting accumulates over eight updates.Source: B. RQ1: Overall Effect Without Sample… · p. 7
  • WiMoT overtakes every baseline from the second stage on WiMANS and Widar3.0, and retains at least 93.8% on each of the first eight Widar3.0 users after the ninth user.Fig. 3 attributes the final differences to retention rather than first-domain accuracy; WiMoT refits OSCL after each update using transported moments from all users with equal weight.Source: B. RQ1: Overall Effect Without Sample… · p. 7, Figure 3 · p. 7

C. RQ2: Plasticity–Stability Balance and the Role of Each Mechanism

WiMoT balances accuracy on newly learned and earlier domains, with its strongest gains in retention; its mechanisms and transport are tested through ablations and robustness settings. Its lead persists across domain orders, rooms, label budgets, and encoders, though the WiMANS results are more mixed across settings.

C. RQ2: Plasticity–Stability Balance and the Role of Each Mechanism

  • WiMoT has the highest Balance on 3DO and Widar3.0 (94.94% and 95.69%), ahead of SI+ER by 3.03 ± 0.53 and 8.51 ± 0.58 pp; on WiMANS its 89.12% is within seed variation of MAS+ER.Its Earlier accuracy is highest on all three streams, while New accuracy stays close to the best baseline; the restricted affine adapter accounts for a small loss in New accuracy.Source: C. RQ2: Plasticity–Stability Balance… · p. 7, Table III · p. 8
  • Without SupCon, AFA falls by 2.54–5.12 pp on all streams; without the historical loss, AFR rises by 0.36–1.11 pp on all streams.Removing historical moments from OSCL lowers AFA by 17.04, 5.31, and 47.10 pp and raises AFR by 30.28, 5.28, and 52.03 pp across the three streams.Source: C. RQ2: Plasticity–Stability Balance… · p. 7, Table IV · p. 9
  • Transport raises AFA from R1 to R3 by 12.05, 3.40, and 31.01 pp across the streams, accounting for most of the full method’s gain over adapting without transport.Without transport, the classifier uses statistics that no longer match test features; the historical loss helps only when the memory is valid.Source: C. RQ2: Plasticity–Stability Balance… · p. 8, Table V · p. 9

D. RQ3: Robustness of the Lead

  • Across six 3DO domain orders, WiMoT leads in every seed by 1.13–6.81 pp; with 25% and 10% of later-domain labels, its leads are 2.81 and 2.40 pp on 3DO and 9.85 and 10.57 pp on Widar3.0.Across five Widar3.0 streams, its AFR ranges from −0.45 to 1.42 pp, while every listed baseline has AFR of at least 3.80 pp.Source: D. RQ3: Robustness of the Lead · p. 8
  • With every tested encoder, WiMoT exceeds the strongest of five baselines on 3DO and Widar3.0 by 3.77–9.56 pp and 13.70–30.99 pp, respectively.On WiMANS it leads with the two MantisV2 encoders, but the strongest baseline is ahead with the five other encoders.Source: D. RQ3: Robustness of the Lead · p. 9
  • Across 42 settings, WiMoT’s AFA lead grows with the strongest baseline’s AFR (Pearson r = 0.94, Spearman ρ = 0.94); after dataset means are removed, r = 0.80 and ρ = 0.62.Across loss-weight ratios from 0.1 to 2, mean AFA varies by at most 0.18, 0.51, and 0.25 pp on 3DO, WiMANS, and Widar3.0.Source: Figure 4 · p. 8, D. RQ3: Robustness of the Lead · p. 9

E. RQ4: Deployment Cost

WiMoT’s deployment workflow encodes new arrivals once, updates from cached basis features, then deletes recordings. Its moment store trades greater storage for retaining aggregate statistics, with compression options and a qualitative privacy rationale.

E. RQ4: Deployment Cost

  • The node encodes a new labeled calibration session once, adds its moments, updates the adapter on cached basis features, refits the classifier, and deletes recordings.The procedure is an analytic deployment design, not a device implementation.Source: E. RQ4: Deployment Cost · p. 9
  • On a home-gateway design, queries use neither stored samples nor domain identity, while the frozen encoder accounts for more than 99.8% of inference computation.Basis, adapter, and classifier add 15.9–16.3 MFLOPs per query; update gradient steps add 0.15–0.68% to the encoding cost of arrivals.Source: E. RQ4: Deployment Cost · p. 9
  • 108.21 to 6.97 MiB on Widar3.0 is the rank-32 moment-memory reduction, with an AFA change of −0.49 to +1.07 pp.Full 512×512 float64 covariances make the store larger than replay arrays; upper-triangle storage halves it without changing predictions.Source: E. RQ4: Deployment Cost · p. 10

A. What the Stored Moments Replace

  • The moment store replaces replay individuals with first and second moments for every individual in each domain and class; removing moments lowers AFA by 5.31–47.10 pp.Adding replay caches to WiMoT changes AFA by +0.17, −0.62, and −0.67 pp; strongest replay baselines forget 2.78–10.25 pp, versus at most 0.58 pp for WiMoT.Source: A. What the Stored Moments Replace · p. 10

B. Privacy

  • Each stored moment aggregates at least 42 recordings and cannot be inverted to recover an individual recording, but small-group aggregates may still reveal information.The authors characterize this protection as qualitative; a formal guarantee would require, for example, noise calibrated to the moments.Source: B. Privacy · p. 11

C. Limitations

WiMoT’s advantages vary with stream length, labeled records, and frozen-feature quality. The authors also identify growing memory and the absence of on-device measurements as deployment boundaries.

C. Limitations

  • WiMANS accuracy advantages depend on order, label budget, and encoder; in three of six domain orders, MAS or SI+ER leads by 0.93–1.95 pp.With 25% and 10% of later-domain labels, MAS is higher by 1.23 ± 2.64 and 1.44 ± 1.46 pp, respectively, within one seed SD.Source: C. Limitations · p. 11
  • WiMoT is less suited to short streams with little baseline forgetting, domains with few labeled records per class, or encoders whose frozen features separate classes poorly.On WiMANS, each class and domain has 42 training records for 512-dimensional moments, so covariance rank is at most 41; the amplitude descriptor carries little class information there.Source: C. Limitations · pp. 11–12
  • Moment memory grows with the number of domains and classes, and latency and energy were not measured on a device.The Widar3.0 partitions separate repetitions rather than recording sessions.Source: C. Limitations · p. 12

VII. CONCLUSION

  • The conclusion reports at most 0.58 pp forgetting and the highest average final accuracy on 3DO and Widar3.0, while naming unannounced shifts, unlabeled arrivals, and device measurements as next steps.On WiMANS, WiMoT matches the strongest replay baseline in accuracy while eliminating its forgetting.Source: §VII · p. 12

A. Classic Baselines

The baseline protocol compares classic continual-learning methods, aligned recent methods, and Wi-Fi-specific approaches on a shared input interface. Replay identities and preprocessing are also standardized or explicitly described.

A. Classic Baselines

  • The classic comparison uses twelve configurations with a shared 9728-dimensional input, while baseline networks remain trainable throughout later updates.Native and replay-augmented regularization coefficients are selected separately on validation AFA, then fixed across datasets, seeds, orders, and Widar3.0 streams.Source: A. Classic Baselines · p. 14

B. Recent Mechanisms with Aligned Historical Individuals

  • Recent methods CN-DPM, DUCT, and DualCP variants are adapted to Mantis features, with replay variants aligned to common reservoir identities and visits.DualCP required adding threshold 0.85 after the prespecified candidates collapsed to a single group on 3DO and WiMANS; both DualCP variants remained far below the strongest baseline on every stream.Source: B. Recent Mechanisms with Aligned… · pp. 14–15

C. Wi-Fi Continual-Learning Methods

  • ConSense and EdgeCL are adapted to the shared 9728-dimensional input; neither receives domain identity at inference.ConSense adds frozen attention and per-domain prefixes, while EdgeCL stores selected exemplars with distilled labels and uses replay during later updates.Source: C. Wi-Fi Continual-Learning Methods · p. 15

A. Feature Normalization

  • Feature preprocessing uses first-domain semantic normalization and block scaling, while WiMANS and Widar3.0 inputs are standardized per record and 3DO preserves amplitude scale.A fixed random amplitude stem and first-domain-fitted PCA shortcut complete preprocessing.Source: A. Feature Normalization · pp. 15–16

B. Common Replay Cache of the Baselines

  • The common replay cache is class-balanced and refreshed after each domain using fixed hash priorities; capacities are 240 windows on 3DO, 36 records on WiMANS, and 42 instances on Widar3.0.The cache is held fixed during the next domain’s training, and evicted examples cannot be recovered.Source: B. Common Replay Cache of the Baselines · p. 16

C. Memory Accounting

WiMoT’s aggregate statistics dominate its memory footprint, while its non-backbone model state remains comparatively small. The statistical head is fitted from moments, with cross-fitting used to select classifier scoring choices.

C. Memory Accounting

  • 512 × 512 float64 class covariances account for 99.8% of WiMoT’s class-moment payload, which occupies 18.04, 54.11, and 108.21 MiB across 3DO, WiMANS, and Widar3.0.Historical second and label-cross moments occupy 3.03, 3.06, and 9.14 MiB, respectively.Source: C. Memory Accounting · p. 16
  • WiMoT updates 82,563/83,337/82,950 parameters on 3DO/WiMANS/Widar3.0 after freezing the first-domain basis, and its non-backbone model tensors occupy approximately 30.08 MiB.These figures exclude the shared frozen backbone, optimizer state, fitted statistical heads, and archival evaluation outputs; aggregates grow with domains and classes.Source: C. Memory Accounting · p. 16

D. Loss Construction

  • The historical loss is a class-balanced squared error of normalized scores against centered one-hot targets, evaluated from stored moments.The score map’s output scale is fixed at the end of the first domain, and each loss term is clipped at zero for numerical safety.Source: D. Loss Construction · p. 16

S-III. OSCL CLASSIFIER

  • The OSCL head is cross-fitted on training data with equal weight per observed task and per observed class within each task, then refitted on full training moments.Cross-fitting selects the score family and gate by held-out-fold log likelihood; only the statistical head is cross-fitted, not the encoder, adapter, or historical-loss map.Source: S-III. OSCL CLASSIFIER · p. 17

A. Additional Seeds and Reproducibility

Additional seeds support WiMoT’s 3DO and WiMANS results, while domain-order tests show stronger order consistency on 3DO than WiMANS. On Widar3.0, WiMoT outperforms the strongest baseline across the main stream and further user orders.

A. Additional Seeds and Reproducibility

  • 93.98 ± 0.15% and 89.57 ± 0.55% AFA: WiMoT leads ConSense and EdgeCL on 3DO and WiMANS across five seeds.A second-machine check reproduced every WiMANS seed and changed 3DO and Widar3.0 AFA by at most 0.09 and 0.07 pp per seed; the package archives code and run records.Source: A. Additional Seeds and Reproducibility · p. 17

B. Domain Orders

  • 1.13–6.81 pp: WiMoT leads in every 3DO order; on WiMANS, it leads in three orders and averages 0.34 pp over SI+ER.WiMANS results vary by order, and WiMoT’s AFR is lowest in four of six orders.Source: B. Domain Orders · pp. 17–18, Table S1 · p. 18

C. Widar3.0: Main Stream, Further User Orders, and Rooms

  • 20.15 ± 1.64 and 20.32 ± 1.83 pp: on Widar3.0, WiMoT exceeds EdgeCL and CN-DPM, respectively, while LwF+ER is the strongest baseline at 82.05 ± 1.07% AFA.The comparison uses a longer stream with seven representation updates after the second domain; Table S2 lists final accuracy per user.Source: C. Widar3.0: Main Stream, Further User… · p. 18, Table S2 · p. 18
  • 94.47 ± 0.45, 94.27 ± 0.38, and 94.79 ± 0.23% AFA: WiMoT leads SI+ER by 9.58–14.52 pp across three further Widar3.0 user orders.WiMoT’s AFR stays at 1.11–1.42 pp, versus 10.33–14.33 pp for SI+ER.Source: C. Widar3.0: Main Stream, Further User… · p. 18

D. Reduced Labels in Later Domains

With fewer labels, WiMoT remains ahead of the strongest baseline on 3DO and Widar3.0, while WiMANS has the lowest forgetting but not the highest AFA. Results also extend across frozen encoders, and rank-32 covariance storage retains performance with a much smaller representation.

D. Reduced Labels in Later Domains

  • +2.81 ± 1.24 and +2.40 ± 0.33 pp: WiMoT beats SI+ER on 3DO at 25% and 10% labels; on Widar3.0 it beats SI+ER and EdgeCL by +9.85 ± 2.12 and +10.57 ± 1.10 pp.On WiMANS, WiMoT has the lowest AFR at both label fractions, while MAS leads AFA by 1.23 ± 2.64 and 1.44 ± 1.46 pp.Source: D. Reduced Labels in Later Domains · p. 19, Table S3 · p. 20

E. Other Frozen Encoders

  • 2.11–4.24 and 6.96–25.03 pp: with six other encoders, the strongest baseline loses more AFA than WiMoT on 3DO and Widar3.0, respectively.WiMoT’s lead is larger than with pretrained MantisV2 on 3DO and five of six encoders on Widar3.0; the strongest baseline also has greater AFR across those encoders.Source: E. Other Frozen Encoders · p. 19, Table S4 · p. 20
  • 89.61%: WiMoT with pretrained MantisV2 has the study’s highest WiMANS AFA, while the strongest baseline with each of five other encoders is ahead in every seed.The study does not separate the effects of coefficients selected with MantisV2 features and the small number of WiMANS training records per class.Source: E. Other Frozen Encoders · p. 19
  • 96.03, 95.56, and 94.82%: randomly initialized MantisV2, TSPulse-r1, and MOMENT-1-large give WiMoT higher 3DO AFA than pretrained MantisV2 at 94.02%.The passage links this pattern to 3DO class information being carried mainly by amplitude statistics.Source: E. Other Frozen Encoders · p. 19

F. Covariance Storage

  • 513r+1 numbers per class: rank-r covariance storage reconstructs each covariance from leading eigenpairs and an isotropic residual; at rank 32, AFA changes by +1.07 ± 0.62 pp on 3DO, is unchanged on WiMANS, and falls by 0.49 ± 0.31 pp on Widar3.0.At rank 8, the Widar3.0 loss is 2.74 ± 0.39 pp.Source: F. Covariance Storage · pp. 19–20

G. Lead and Baseline Forgetting

Across 42 settings, WiMoT’s lead is associated with baseline forgetting, but the analysis does not identify a mechanism. The input-feature analysis specifies comparisons across Mantis, amplitude, and combined inputs, with controlled recipes and shared run conditions.

G. Lead and Baseline Forgetting

  • 0.94: WiMoT’s lead correlates with the strongest baseline’s AFR across 42 settings, but the analysis says this correlation does not identify a mechanism.Within-dataset Spearman correlations are 0.40 on 3DO, 0.59 on WiMANS, and 0.75 on Widar3.0; the p-values are descriptive because settings share recordings, seeds, and baselines.Source: G. Lead and Baseline Forgetting · pp. 20–21, Table S7 · p. 21

S-V. INPUT FEATURES OF WIMOT AND THE BASELINES

  • The input study varies four baselines across Mantis-only, combined, and amplitude-only features, then repeats all 21 baselines with Mantis features alone.Inputs use specified normalization and coefficient-selection procedures; runs use seeds 42–44, fixed partitions, domain orders, replay caches, and schedules.Source: S-V. INPUT FEATURES OF WIMOT AND THE… · p. 21, A. Setting · p. 21

A. Setting

  • Two amplitude-only WiMoT controls preserve the main recipe and classifier but alter whether the semantic stem and PCA shortcut read the descriptor or only the fixed random stem is used.They are analysis controls; the main configuration remains unchanged.Source: A. Setting · pp. 21–22

B. Adding the Amplitude Descriptor

Adding the amplitude descriptor to the Mantis block leaves the strongest-baseline comparison largely intact, with WiMoT ahead across all three datasets. Amplitude-only results vary by dataset, and comparisons using Mantis alone preserve positive leads on 3DO and Widar3.0.

C. Amplitude Descriptor Alone

  • 99.02 ± 0.16%: the amplitude-only WiMoT variant whose basis reads the descriptor beats SI+ER, CN-DPM, and DER++ on 3DO, while the main configuration is 5.00 ± 0.10 pp lower.With amplitude-only input, those three baselines exceed the main WiMoT configuration.Source: C. Amplitude Descriptor Alone · p. 22
  • 23.97–32.51%: all methods remain in this WiMANS AFA range with amplitude-only input, indicating that the descriptor carries little class information there.The joint reference reaches 28.19%, compared with 11.1% chance level for nine classes.Source: C. Amplitude Descriptor Alone · p. 22
  • 94.00 ± 0.15%: the amplitude-basis WiMoT variant on Widar3.0 exceeds SI+ER, DER++, EdgeCL, and CN-DPM by 41.46–66.64 pp in every seed.The main configuration remains higher than the variant without a foundation encoder by 1.16 ± 1.08 pp in every seed.Source: C. Amplitude Descriptor Alone · pp. 22–23

D. Mantis Block Alone for All Baselines

A. Source and Partitions

  • 450, 150, and 150 instances per user: the Widar3.0 training, validation, and test partitions separate repetitions within a session rather than recording sessions.Room 2 has two users and Room 3 has four; their streams use the same partitions, features, recipes, and coefficients as Room 1.Source: A. Source and Partitions · p. 23, C. Further Rooms · p. 23

B. Settings and Budgets

  • 42 training instances: the shared Widar3.0 replay cache holds seven examples per class, or about 8.54–8.61% of each user’s training arrivals.The reported recipes and coefficients were selected on 3DO and WiMANS and kept unchanged for Widar3.0.Source: B. Settings and Budgets · p. 23

S-VII. PLASTICITY AND STABILITY IN THE ABLATION

The ablations separate the roles of contrastive learning, historical moments, and adapter updates; classifier replacement favors OSCL on two streams, while loss-weight sweeps reveal forgetting sensitivity.

S-VII. PLASTICITY AND STABILITY IN THE ABLATION

  • Removing SupCon lowers new-domain accuracy by 5.60, 6.85, and 6.45 pp without raising AFR; removing Lhist raises AFR by 0.36, 1.11, and 0.68 pp.Without stored historical moments, earlier-domain accuracy falls by 30.02, 7.87, and 52.73 pp; freezing the adapter lowers new-domain accuracy by 0.43, 3.52, and 6.90 pp.Source: S-VII. PLASTICITY AND STABILITY IN THE… · p. 24, Table S10 · p. 25

S-VIII. CLASSIFIER REPLACEMENT

  • OSCL has higher AFA than every alternative on 3DO and Widar3.0 in all five seeds, leading the strongest LDA and QDA alternatives by 1.93±0.61 and 3.84±0.96 pp.On WiMANS, OSCL leads nearest class mean by 1.73 ± 0.77 pp and QDA by 8.64 ± 1.09 pp, but is not separated from LDA, which is lower by 0.74±1.39 pp.Source: S-VIII. CLASSIFIER REPLACEMENT · pp. 24–25, Table S11 · p. 25

S-IX. SENSITIVITY TO THE LOSS WEIGHTS

  • AFA is insensitive for loss-weight ratios between 0.1 and 2, but on Widar3.0 AFR rises from 0.31 pp at λh = 2 to 1.33 pp at λh = 0.Removing either weighted loss lowers AFA on WiMANS and Widar3.0 but not 3DO; removing SupCon from both phases further lowers AFA on all three streams.Source: S-IX. SENSITIVITY TO THE LOSS WEIGHTS · p. 25, Table S12 · p. 26

Related papers

Loading 2610.32660v2…