acceptodds
Under review as a conference paper at ICLR 2027

LATF-AD: Learning Locally Adaptive Time-Frequency Normality for Time Series Anomaly Detection

Abstract

Time series anomaly detection (TSAD) is essential for reliable monitoring of complex systems. Although frequency-aware modeling can reveal latent anomalies, global spectral representations often obscure localized transients, while a single fixed transformation inadequately captures heterogeneous temporal dynamics across varying resolutions. To address these limitations, we propose LATF-AD, a novel framework for locally adaptive time-frequency normality modeling. LATF-AD constructs an ordered bank of heterogeneous time-frequency views and learns content-dependent, locally adaptive routing to accommodate dynamic spectral patterns via a weighted consensus. It further feeds the consensus into dual prediction and reconstruction heads, while incorporating a prototype-guided raw-domain branch to modulate anomaly scoring against time-frequency artifacts. Extensive experiments on the comprehensive TSB-AD Benchmark demonstrate that LATF-AD consistently achieves state-of-the-art performance across both point-wise and range-based metrics.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Anomaly Detection, Time Frequency Analysis.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fC1EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LATF-AD combines four complementary time-frequency views: short-window STFT, long-window STFT, continuous wavelet transform, smoothed pseudo Wigner–Ville distribution through a local content-dependent router, learns reconstruction and prediction heads, and incorporates raw-patch prototype evidence through a calibrated multiplicative-additive score. Experiments on TSB-AD-U/M report promising improvements and include component ablations.

Critical strengths
  1. The heterogeneous-view motivation is reasonable, and content-dependent fusion is relevant to anomalies with different temporal and spectral structures.
  2. The paper covers both univariate and multivariate settings and several metrics. Tables 2 and 6 provide useful view, branch, and head ablations.
Critical weaknesses

1.The prediction head is not causal as claimed in Eq. (5). Its GroupNorm aggregates across time, allowing future inputs to affect earlier predictions, as confirmed by a suffix-perturbation test. This is acceptable for offline detection, but conflicts with the stated forecasting mechanism. Please clarify the intended information flow and revise the formulation accordingly.

2. The central benefit of local adaptation needs a more direct control. Uniform fusion and view removal are useful, but a learned globally static router with the same encoders and heads would better isolate the value of locality. A raw-only result and simple additive fusion would clarify the benefit of the proposed score. Figure 5 already includes individual zero-coefficient cases; my request is for focused attribution, not an entirely new evaluation suite. The existing ablation scripts should also be supplied so these comparisons can be checked.

3. The uncertainty reported in Table 4 remains unexplained. The paper states that experiments are repeated ten times but does not define “±”. Genesis contains one series and has VUS-ROC 0.9991 ± 0.0179. For ten observations in [0,1] with mean m, their sample standard deviation cannot exceed sqrt(10)(1−m): approximately 0.00285 here, thus 0.0179 cannot be that standard deviation. The supplied reproduce.py runs a single configured seed and aggregates file means; it contains neither the ten-run results nor the uncertainty/table-generation procedure. This does not establish that the experiments were not performed, but the raw values, meaning of the error term, and aggregation procedure are necessary to assess the evidence.

Questions for authors
  1. How is the causal prediction claim in Eq. (5) consistent with GroupNorm aggregating across time and allowing future information to affect earlier predictions?

  2. Does local routing outperform learned static weights under the same backbone? How does the proposed fusion compare with raw-only scoring and simple additive fusion?

  3. What does Table 4's “±” denote? Can you provide the ten Genesis scores and the table-generation procedure?

  4. Since S=TFS=TF when g=0g=0, what evidence supports the claimed false-alarm control? How does patch normalization affect the raw branch’s ability to detect level and scale changes?

3 comments

Thank you for the insightful comments!


Q1: GroupNorm and the causal prediction claim in Eq. (5).

A1:

GroupNorm computes statistics across time, including in evaluation mode. Consequently, the one-frame input shift and left-padded convolutions do not make the submitted temporal branch strictly causal. The symmetric encoders and centered routing operations also introduce future dependence.

We tested 36 trained checkpoints from 12 series selected in advance using input metadata and three seeds. We zeroed either the latent suffix zℓ:Lz_{\ell:L} before the one-frame shift or the complete TF network's input suffix M:,:,ℓ:LM_{:,:,\ell:L}, then compared the prediction at frame ℓ\ell with its unmodified value. Model parameters and fitted preprocessing statistics were fixed. The maximum absolute changes on CPU in evaluation mode were as follows.

Perturbed inputMaximum absolute change at frame ℓ\ell
Latent sequence before the one-frame shift0.827941
TF-frame inputs to the complete TF network2.586952

These results confirm the submitted implementation's future dependence. We will describe LATF-AD as an offline anomaly detector and revise Eq. (5) to M^ℓpred=[Pψpred(S(z1:L))]ℓ\widehat{M}^{\mathrm{pred}}_{\ell}=[P_{\psi_{\mathrm{pred}}}(\mathcal{S}(z_{1:L}))]_{\ell}, where S(z1:L)=(0,z1,…,zL−1)\mathcal{S}(z_{1:L})=(0,z_1,\ldots,z_{L-1}). The temporal branch learns temporal consistency from shifted latent representations. Fitting and calibration use the benchmark-designated training prefix, while inference uses the complete observed sequence. These clarifications retain the submitted implementation.


Q2: Learned static routing, Raw-only scoring, additive fusion, and implementation details.

A2:

The submitted Table 2 reports the following ablation results.

VariantTSB-AD-U VUS-PRTSB-AD-M VUS-PR
Full LATF-AD0.600.46
Uniform TF fusion0.560.43
Without Raw, TF-only0.560.44

These comparisons support the contribution of learned fusion relative to equal weighting and the contribution of Raw evidence to the complete detector.

The submitted score is St=TFt(1+αgt)+βgtS_t=TF_t(1+\alpha g_t)+\beta g_t, with α=0.25\alpha=0.25 and β=20\beta=20. It contains TF evidence, a joint term αTFtgt\alpha TF_tg_t, and an additive Raw term βgt\beta g_t. The joint term increases Raw's contribution when TF evidence is also strong, while the additive term allows Raw evidence to contribute when TF is weak. Algebraically, α=0\alpha=0 yields additive fusion, β=0\beta=0 retains multiplicative modulation, and α=β=0\alpha=\beta=0 gives TF-only scoring. These identities explain the score's components. The empirical claims above are based on the reported uniform-fusion and Raw-removal ablations.


Q3: Meaning of Table 4's uncertainty and the ten Genesis scores.

A3:

The reported Genesis entry, 0.9991 ± 0.0179, is incorrect. We completed a new ten-seed evaluation using the submitted implementation and fixed benchmark JSON configurations. It covers all 350 U and 180 M evaluation files, with seeds 20260805 through 20260814, totaling 5,300 file-seed runs. Both TF and Raw were retrained for every run.

Here, ± denotes the sample standard deviation across ten seed-level means. For each subset or benchmark, Qs=1N∑i=1Nqi,sQ_s=\frac{1}{N}\sum_{i=1}^{N}q_{i,s}, Qˉ=110∑s=110Qs\bar{Q}=\frac{1}{10}\sum_{s=1}^{10}Q_s, and SD=19∑s=110(Qs−Qˉ)2\mathrm{SD}=\sqrt{\frac{1}{9}\sum_{s=1}^{10}(Q_s-\bar{Q})^2}. Files have equal weight within each seed. Genesis contains one file, so its ten seed-level values are the individual run scores.

SeedNewly measured Genesis VUS-ROC
202608050.9993083372
202608060.9994086861
202608070.9996778981
202608080.9994784622
202608090.9995228617
202608100.9994514231
202608110.9997854757
202608120.9995231806
202608130.9994223434
202608140.9997546919

These runs give Genesis VUS-ROC of 0.999533 ± 0.000157 and VUS-PR of 0.828734 ± 0.049034. For SED's two files, the same aggregation gives VUS-ROC of 0.999215 ± 0.000057 and VUS-PR of 0.977010 ± 0.002599.

The complete new benchmark results are as follows.

MetricTSB-AD-U mean ± SDTSB-AD-M mean ± SD
VUS-PR0.605145 ± 0.0048150.450707 ± 0.002832
VUS-ROC0.929288 ± 0.0007950.796448 ± 0.001311
Range-F10.522265 ± 0.0043230.437714 ± 0.004507
AUC-PR0.509642 ± 0.0063120.407660 ± 0.002735
AUC-ROC0.909153 ± 0.0007460.773082 ± 0.001306
Point-F10.565629 ± 0.0054150.458147 ± 0.001765

These are newly measured results. We will use them to correct the LATF-AD row in Table 1, Table 4's means and uncertainty values, and the corresponding performance statements. An independent recomputation matched all 504 mean and SD entries across 40 dataset subsets and two benchmark rows within an absolute difference below 10−1210^{-12}.

Q4: False-alarm control and normalization of level and scale changes.

A4:

For score fusion, the submitted nonnegative evidence gives St−TFt=gt(0.25TFt+20)≥0S_t-TF_t=g_t(0.25TF_t+20)\geq0. Raw therefore cannot cancel a TF alert at the same numerical threshold. Its contribution is complementary evidence enhancement. In the submitted Table 2, removing Raw reduces VUS-PR from 0.60 to 0.56 on U and from 0.46 to 0.44 on M. We will revise the artifact-rejection explanation to match this scoring behavior.

For each channel within a patch, Raw computes IN⁡(x)=x−μ(x)Var⁡(x)+10−5\operatorname{IN}(x)=\frac{x-\mu(x)}{\sqrt{\operatorname{Var}(x)+10^{-5}}}. Thus, IN⁡(x+c)=IN⁡(x)\operatorname{IN}(x+c)=\operatorname{IN}(x) for a constant offset cc. For positive scaling, IN⁡(ax)=x−μ(x)Var⁡(x)+10−5/a2\operatorname{IN}(ax)=\frac{x-\mu(x)}{\sqrt{\operatorname{Var}(x)+10^{-5}/a^2}}, where a>0a>0. Whole-patch offsets are removed, and positive scaling is approximately removed when the variance dominates the stabilizer.

Patches crossing the onset or end of a level or scale change can retain a within-patch shape difference because they contain both regimes. Patches wholly inside a sustained change lose the offset or much of the scale evidence. The TF branch uses prefix-fitted normalization, so it does not share Raw's per-patch invariance. We will clarify these normalization properties and the distinction between anomaly boundaries and interiors.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer,
Thank you for your time and constructive feedback, which have helped improve the paper.
We would greatly appreciate it if you could review our responses and share any remaining concerns when you have a chance. We would be happy to provide further clarification or answer any additional questions.
We look forward to your further feedback. Thanks again!

Reviewer_2ewgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a unsupervised time series anomaly detection framework, LATF-AD, which focuses on the issue that a single time-frequency transform often struggles to provide both high temporal resolution and spectral resolution. This method utilizes short-window STFT, long-window STFT, CWT, and WVD to construct four heterogeneous views. A content-dependent router assigns timestamps-specific weights to these views, and the resulting consensus representation is then processed by the reconstruction head and prediction head. Simultaneously, an original time-domain patch encoder constructs a prototype memory from normal training prefixes. The time-frequency residuals and prototype distances undergo robust calibration and combination, ultimately yielding an anomaly score. This method was evaluated on both univariate and multivariate TSB-AD benchmarks, in comparison with a series of time-domain, frequency-domain, and baseline models. LATF-AD reported superior comprehensive performance across six point-wise and interval metrics, with a VUS-PR score of approximately 0.60 on TSB-AD-U and a score of approximately 0.46 on TSB-AD-M. Ablation experiments suggest that multiple views, dynamic fusion, a dual-head structure, and the original patch branch may all contribute to the final performance to some extent.

Critical strengths

1、Clear central motivation. The paper points out an important limitation of Time Series Anomaly Detection (TSAD) in frequency perception: no single time-frequency resolution can universally apply to short-term transients, long-term rhythmic deviations, and non-stationary behaviors. The proposed idea of routing between heterogeneous time-frequency views at each time position is directly related to this limitation. 2、The experimental results are extensive and show promising outcomes. LATF-AD was evaluated on TSB-AD-U and TSB-AD-M using six point-wise and interval metrics, as well as a fairly extensive set of recent baselines. This improvement is particularly meaningful on the primary VUS-PR metric of TSB-AD-U, with the reported score increasing from 0.54 for the strongest baseline to 0.60. The authors also provided ablation experiments for single-view, leave-one-out, uniform fusion, dual-head, and original branches, accompanied by results and computational overviews for each dataset.

Critical weaknesses

1、The main control experiment, uniform fusion was used instead of dynamic routing. This only indicates that the learned router is superior to assigning equal weights to each view, but it does not establish that timestamp-level adaptation is the reason for performance improvement. Figure 1 shows that weights change, but it does not indicate that the selected high-weight view is indeed the most reliable at that timestamp. 2、The ablation experiment only indicates that adding a second detector can improve performance, but it cannot prove that it serves as an independent verification mechanism for rejecting spectral artifacts. 3、The exact definition of Rob(⋅;⋅), the final values of a,b,α,β, and the tuning and implementation protocol for the baselines are not sufficiently explicit. These details matter because several reported margins on TSB-AD-M are only 0.01- 0.03.

Questions for authors

1、Could the authors compare the current timestamp-dependent router against: learned but time-invariant global view weights, the temporal average of the learned routing weights frozen at inference, and temporally shuffled routing weights, while keeping the encoders, heads, parameter budget, and scoring function fixed ? 2、Since Eq.9 satisfies St≥TFt for non-negative gt, how can the raw-domain branch suppress false positives caused by time-frequency artifacts ? 3、Please provide the exact definition of Rob(⋅;⋅), the selected values and tuning ranges of a,b,α,β, the precise TSB-AD tuning and evaluation file lists used for the main table, and the source and tuning procedure for every baseline.

2 comments

Thank you for the insightful comments!


Q1: Timestamp-dependent routing versus static, averaged, and shuffled weights.

A1:

The submitted ablation provides evidence for learned fusion relative to equal weighting. In Table 2, replacing the learned router with uniform TF-view weights reduces VUS-PR from 0.60 to 0.56 on TSB-AD-U and from 0.46 to 0.43 on TSB-AD-M.

The router uses frame-dependent view statistics to allocate weights. Two auxiliary objectives guide this allocation. The first constructs a target from each view's squared temporal innovation, normalized by its mean innovation within the training context, and applies cross-entropy to the clean routing weights with coefficient 0.1. The second adds Gaussian noise to one randomly selected TF view and penalizes the weight assigned to that view with coefficient 0.01. Thus, the allocation is trained using relative temporal variation and known artificial corruption. Figure 1 illustrates the learned allocations over time. We will describe these weights as view allocations and tie the empirical conclusion specifically to the learned-versus-uniform comparison in Table 2.


Q2: Can the Raw branch suppress false positives caused by time-frequency artifacts?

A2:

The Raw branch contributes complementary anomaly evidence. Both submitted configurations use nonnegative TFtTF_t and gtg_t, with α=0.25\alpha=0.25 and β=20\beta=20. Therefore, St=TFt(1+0.25gt)+20gtS_t=TF_t(1+0.25g_t)+20g_t and St−TFt=gt(0.25TFt+20)≥0S_t-TF_t=g_t(0.25TF_t+20)\geq0. At the same numerical threshold, Raw cannot cancel an alert already triggered by TF. When gt=0g_t=0, the TF score is unchanged.

The multiplicative term reinforces evidence when both branches respond, while the additive term contributes when Raw responds and TF evidence is weak. In the submitted Table 2, removing Raw reduces VUS-PR from 0.60 to 0.56 on TSB-AD-U and from 0.46 to 0.44 on TSB-AD-M. These results support Raw's contribution to detection performance. We will revise the manuscript to explain its role as complementary raw-domain evidence enhancement. The ablation concerns VUS-PR and does not by itself establish a lower false-positive rate.


Q3: Calibration, coefficients, file lists, and evaluation protocol.

A3:

For reference values RR, the submitted calibration is defined by mR=median⁡(R)m_R=\operatorname{median}(R), sR=max⁡(1.4826median⁡r∈R∣r−mR∣,10−5)s_R=\max(1.4826\operatorname{median}_{r\in R}|r-m_R|,10^{-5}), and Rob⁡(x;R)=max⁡((x−mR)/sR,0)\operatorname{Rob}(x;R)=\max((x-m_R)/s_R,0).

Temporal and reconstruction evidence are calibrated separately and combined as qt=aRob⁡(ettemp;Rtemp)+bRob⁡(etrec;Rrec)q_t=a\operatorname{Rob}(e_t^{\mathrm{temp}};R_{\mathrm{temp}})+b\operatorname{Rob}(e_t^{\mathrm{rec}};R_{\mathrm{rec}}). The combined evidence is calibrated again as TFt=Rob⁡(qt;Rq)TF_t=\operatorname{Rob}(q_t;R_q). Raw evidence is gt=min⁡(Rob⁡(etraw;Rraw),10)g_t=\min(\operatorname{Rob}(e_t^{\mathrm{raw}};R_{\mathrm{raw}}),10). Each reference is fitted on the corresponding evidence from the separately processed benchmark-designated training prefix.

The submitted JSON configurations use the following fixed values. The same configurations were used in the completed ten-seed evaluation.

Benchmarkaabbα\alphaβ\betaRouter temperature
TSB-AD-U3.210.25200.05
TSB-AD-M12.80.250.25200.15

The two temperatures differ in the submitted configurations, and we will correct the manuscript accordingly.

We checked file membership against the official lists at commit 303dffbfc11612995b17b06633f8c9d0ea0e38e5. These contain 48 U tuning files, 350 U evaluation files, 20 M tuning files, and 180 M evaluation files. All 598 files passed presence, finite-value, and binary-label checks. The completed repeated evaluation covers all 530 evaluation files and retains the benchmark-designated training-prefix boundaries without label-based filtering.

Evaluation uses the complete sequence. The maximum VUS buffer is derived from the first channel's autocorrelation, and VUS uses 250 thresholds. Point-F1 uses the precision-recall threshold sweep, while Range-F1 uses the official 100-threshold sweep. These F1 scores use benchmark oracle thresholds. The evaluator matched the checked official implementations in 156 comparisons on synthetic and data-derived cases and 72 further comparisons on six saved score arrays.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer,
Thank you for your time and constructive feedback, which have helped improve the paper.
We would greatly appreciate it if you could review our responses and share any remaining concerns when you have a chance. We would be happy to provide further clarification or answer any additional questions.
We look forward to your further feedback. Thanks again!

Reviewer_DBGqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a locally adaptive time-frequency unsupervised temporal anomaly detection framework, LATF-AD. It utilizes various time-frequency transformations to construct multiple views and employs a content-aware routing network and a temporal prototype branch to fuse and calibrate residuals to calculate anomaly scores. On the TSB-AD-U/M dataset, the results outperform the baseline, and ablation experiments confirm the module's effectiveness. While the topic is practical, some descriptions differ from the implementation, and there is insufficient comparison with state-of-the-art (SOTA) findings.

Critical strengths

The coherent multi-resolution architecture addresses heterogeneous anomalies, achieving dynamic representation through multiple time-frequency views and adaptive routing to adapt to various monitoring scenarios.

This work simultaneously tests on univariate and multivariate datasets, providing comprehensive metrics, thorough ablation and overhead experiments, stable results, and open-source code.

Critical weaknesses

There are discrepancies between the method implementation and its description: symmetric convolution and temporal GroupNorm violate causal constraints, and SPWVD lacks temporal smoothing; the prototype branch negative sampling scheme is simple, and the comprehensive scoring is insufficient to filter out unfounded frequency domain false anomalies, affecting the explanatory power of the mechanism, although the framework can be used for offline detection.

Without comparison with state-of-the-art models on the TSB-AD leaderboard, the state-of-the-art conclusions are not convincing; the reliability of the route weight matching view cannot be verified, the original scores and reproduction materials are incomplete, and it is difficult to independently verify the adaptive routing gain.

Questions for authors

Q1. How to reconcile the causal prediction claim with actual code implementation?

Please verify causality via zero-out test: whether modifying inputs from frame l onward changes prediction at frame l. If strict causality is not intended, revise the formulation accordingly. Also explain the SPWVD implementation discrepancy on temporal smoothing. A strictly causal encoder experiment will affect my assessment.

Q2. Can you update benchmark comparison with current top TSB-AD methods?

Please compare with CHARM and other strong methods under strictly matched protocols (data split, preprocessing, threshold rule). Provide per-series paired results and confidence intervals to clarify the real performance ranking.

Q3. Can you provide stronger evidence for adaptive routing mechanism and raw-branch effect?

Compare adaptive routing with parameter-matched static fusion and alternative gating methods. Add anomaly-type-stratified analysis on routing weights. Clarify the negative sampling strategy, and test whether instance normalization limits level-shift detection.

9 comments

Thank you for the insightful comments!


Q1: Causal prediction and SPWVD implementation.

A1:

We agree that the causal constraint in Eq. (5) is not supported by the submitted implementation. Although the temporal branch shifts its latent input by one frame and uses left-padded convolutions, GroupNorm aggregates statistics across time, including during evaluation. The symmetric view encoders and centered routing operations introduce additional future dependence.

We performed the requested suffix-zeroing test on 36 trained checkpoints of the submitted architecture, covering 12 series selected in advance using input metadata and three training seeds. We zeroed either the latent sequence zℓ:Lz_{\ell:L} before the one-frame shift or the TF-frame inputs M:,:,ℓ:LM_{:,:,\ell:L}, and compared the prediction at frame ℓ\ell with its unmodified value. The table reports the maximum absolute change across the tested conditions on CPU, in evaluation mode, with model parameters and fitted preprocessing statistics fixed.

Perturbed inputMaximum absolute change in the prediction at the cut frame
Latent sequence before the one-frame shift0.827941
TF-frame inputs to the complete TF network2.586952

These results confirm that the submitted network does not satisfy the stated causal constraint. We will explicitly describe LATF-AD as an offline anomaly detector and interpret the temporal branch as learning temporal consistency from shifted latent representations. Specifically, Eq. (5) will be revised to M^ℓpred=[Pψpred(S(z1:L))]ℓ\widehat{M}^{\mathrm{pred}}_{\ell}=[P_{\psi_{\mathrm{pred}}}(\mathcal{S}(z_{1:L}))]_{\ell}, where S(z1:L)=(0,z1,…,zL−1)\mathcal{S}(z_{1:L})=(0,z_1,\ldots,z_{L-1}).

This formulation describes the implemented computation without restricting its dependence to past frames. We will revise the corresponding causal and forecasting statements throughout the manuscript. Fitting and calibration use the normal training prefix, while inference uses the complete observed sequence. The centered time-frequency transforms also use future raw samples. We acknowledge the value of a strictly causal architectural comparison; the manuscript correction alone does not establish performance under strict causal constraints.

Regarding SPWVD, the submitted implementation constructs analytic-signal lag correlations, convolves them along the lag axis with a normalized seven-point Hann kernel, and then applies FFT magnitude, log compression, and resizing. It does not perform independent smoothing across neighboring center times. The lag convolution also differs from the multiplicative lag window in Eq. (11). We will therefore rename this feature a WVD-derived magnitude view, replace Eq. (11) with the implemented definition, and remove the description of independent temporal smoothing and claims that rely on it. These revisions align the manuscript with the submitted implementation.


Q2: Current strong TSB-AD methods.

A2:

Our manuscript was submitted on September 26, 2026. The official leaderboard entries for CHARM(kNN), CHARM(ZS), and Syntropic were added on October 20, 2026, after submission, as documented in the leaderboard update.

We distinguish these updated CHARM entries from the earlier CHARM baseline, which was already listed before submission. The updated implementation was available in a public pull request before submission, and its September 15 README reported results while stating that deployment of the corresponding model endpoint remained outstanding. Syntropic's benchmark pull request was opened on October 11, after our submission.

CHARM also has a different model-access setting. Its official model card states that the pretrained weights and training/replication code are not publicly released. The public SDK and benchmark scripts access the model through a hosted API. These facts clarify availability and independent-reproduction constraints.

LATF-AD is trained independently on each series' normal prefix without external pretraining. We will explicitly distinguish this setting from methods that use pretrained representations.

Q3: Routing reliability, negative sampling, and instance normalization.

A3:

1. Adaptive routing.

The submitted ablation results in Table 2 show that replacing learned routing with uniform fusion reduces VUS-PR from 0.60 to 0.56 on TSB-AD-U and from 0.46 to 0.43 on TSB-AD-M. This comparison supports the benefit of the submitted learned fusion rule relative to equal weighting.

The routing weights depend on each view's mean, energy, frame-to-frame variation, locally averaged energy, cross-view energy discrepancy, and a learned view embedding. Two auxiliary objectives train this allocation. The first constructs a target from squared temporal innovation, normalized by each view's mean innovation within the training context, and applies cross-entropy to the clean routing weights with coefficient 0.1. The second adds Gaussian noise to one randomly selected view and penalizes its routing weight with coefficient 0.01. These objectives train the router to respond to relative temporal activity and discount artificially corrupted views. Figure 1 illustrates how the resulting allocation changes across temporal regimes. We will make this operational meaning of the routing weights explicit in the manuscript.

2. Contribution of the Raw branch.

In the submitted Table 2, removing the Raw branch reduces VUS-PR from 0.60 to 0.56 on TSB-AD-U and from 0.46 to 0.44 on TSB-AD-M. The branch supplies complementary evidence by comparing raw-patch embeddings with normal prototypes. In Eq. (9), the interaction term amplifies TF evidence when Raw evidence is also present, while the additive term allows Raw evidence to contribute when TF evidence is weak. When gt=0g_t=0, St=TFtS_t=TF_t. We will describe this mechanism as complementary evidence enhancement and clarify the corresponding explanation in the manuscript.

3. Negative-sampling strategy.

Positive patches start one or two samples before or after the anchor, providing highly overlapping local pairs. Let qiq_i and qi+q_i^+ denote unit-normalized projected anchor and positive embeddings. The implementation uses di+=1−qi⊤qi+d_i^+=1-q_i^\top q_i^+ and di−=max⁡j≠i(1−qi⊤qj+)d_i^-=\max_{j\ne i}(1-q_i^\top q_j^+), with Ltriplet=110B∑i=1Bmax⁡(di+−di−+0.1,0)\mathcal{L}_{\mathrm{triplet}}=\frac{1}{10B}\sum_{i=1}^{B}\max(d_i^+-d_i^-+0.1,0).

The negative is selected from the other positive patches in the minibatch, excluding the paired diagonal. Maximizing cosine distance selects the easiest available negative. The candidate pool does not enforce temporal separation, so patches may overlap. The Raw encoder is also trained with an auxiliary adjacency-discrimination objective during the early training phase. We will specify these sampling and training details explicitly, including the distinction between distance in embedding space and separation in time.

4. Instance normalization and level shifts.

Patch instance normalization emphasizes within-patch temporal shape. For each channel, it computes IN⁡(x)=x−μ(x)Var⁡(x)+10−5\operatorname{IN}(x)=\frac{x-\mu(x)}{\sqrt{\operatorname{Var}(x)+10^{-5}}}. Consequently, a constant offset applied to an entire patch is removed. For positive scaling, IN⁡(ax)=x−μ(x)Var⁡(x)+10−5/a2\operatorname{IN}(ax)=\frac{x-\mu(x)}{\sqrt{\operatorname{Var}(x)+10^{-5}/a^2}}, where a>0a>0, giving approximate scale invariance when the variance dominates the stabilizer.

The effect on level-shift detection depends on the patch position. Patches crossing the onset or end of a shift contain a within-patch transition that can remain visible after normalization. Patches wholly within a uniformly shifted interval lose the constant-offset evidence, making the Raw branch less sensitive to that component. The TF branch instead uses prefix-fitted normalization, so this patch-wise offset invariance is not imposed on the complete detector. We will clarify this distinction between boundary transitions, shifted interiors, and the two branches in the manuscript.

The authors clarified in their response: The model has a non-causal architecture, and the prediction head incorporates future information; it should be considered an offline anomaly detection method, and the related method description needs to be revised. The SPWVD implementation does not match the original formula and cannot be explained using the original theory. Please confirm that all experiments use this implementation and distinguish between valid empirical conclusions and invalid theoretical interpretations. The original criticisms are valid; simply modifying the description is insufficient to eliminate the impact on the methodological conclusions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.