acceptodds
Under review as a conference paper at ICLR 2027

Perceive Fast, Think Slow: Adaptive Periodic Tokenization for Time Series Analysis

30 views

Abstract

Patch-based transformers encode raw observations into fixed patches and then rely on inter-patch attention for both dependency modeling and prediction; the same token grid must serve both purposes. PeCo-TS separates these two purposes into coexisting pathways. The Fast Path produces full-resolution timestep features with linear attention, while a spectral segmenter pools them into period-guided segment tokens for scale-adaptive dependency modeling. Temporal reprojection fuses both representations before the output head, so the prediction head never operates on compressed tokens alone. The complete design achieves best or tied-best forecasting MSE on five of seven datasets and the highest displayed PA-F1 in the MSL pathway control. The pathway controls reveal an asymmetric pattern: retaining only the full-resolution pathway trails the full model by a small margin, while retaining only the segment pathway falls far behind. This indicates that aggregation and fine-grained detail are complements, not substitutes. The results show that separating prediction-scale perception from segment-scale conceptualization—rather than forcing both purposes into the same token grid—is a viable design principle for time-series transformers.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Multivariate Time Series, Time-Series Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pFdnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PeCo-TS in that author argues that in patch-based Transformers a single token grid must serve two roles at once: carrying information for inter-token dependency modelling and carrying fine-grained detail for the prediction head. The paper splits these roles into two coexisting pathways. A Fast Path runs per-channel linear attention over individual timesteps. A spectral segmenter estimates one effective period per sample and channel from a softmax-weighted spectral centroid (T = L / keff), builds a cosine comb with that period, and converts it to hard boundaries via NMS; Fast Path features are mean-pooled within the resulting segments into variable-count tokens. A Slow Path applies dual-axis attention (DA2) over segment ordinal and channel, mixed by a layer-wise scalar gate. Temporal reprojection spreads segment context back to timesteps with a Gaussian kernel and the result is added to the unpooled Fast Path features before task heads. Emprical results are shown on 4 tasks heads viz. forecasting, classification, imputation and anomaly detection.

Critical strengths

S1. The core idea is simple and well motivated. Using variable-span segment tokens for dependency modeling while keeping an unpooled timestep stream for the prediction head is a reasonable answer to the patch-size trade-off in PatchTST-style models. Unlike TimesNet/TimeMixer++, the span is chosen per sample and per channel rather than shared across the batch, which is a real difference from prior period-based methods.

S2. The appendix is thorough. It includes boundary perturbatio, detector sensitivity with ANOVA, reprojection width, Fast Path width vs. latency, missing observations, and an FFT vs. Lomb–Scargle comparison.

S3. The method is evaluated on four task families (forecasting, classification, imputation, anomaly detection), not just forecasting. The short derivation in Appendix F.1 (Eq. 12) is also a nice touch: it shows clearly that attention over pooled tokens weights by segment membership instead of timestep count.

Critical weaknesses

W1. Missing and weak baselines. This is my major concern. TimeMixer++ is discussed in related work but not compared, even though it uses FFT periods and dual-axis attention on the same four tasks. Recent strong forecasters such as OLinear and Time-O1 are also missing. Some reproduced baselines are much worse than their published L=96 numbers:

  • iTransformer: Electricity 0.224 vs. 0.178 published, Traffic 0.612 vs. 0.428.
  • TimesNet: Traffic 0.903 vs. 0.620. With the published numbers, PeCo-TS would no longer win on Electricity or Traffic. Against TimeMixer++ / OLinear, PeCo-TS trails by roughly 18–26% on Electricity and Traffic and 8–15% on Weather. Suggested comparison: Include 'strongest' baselines while comparying such as Timemixer++(ICLR 2025), MSTN (TMLR 2026), OLinear (Yue et al., NeurIPS 2025), Time-o1 (Wang et al., NeurIPS 2025).

W2. No variance and several internal inconsistencies. All results are single runs, while the claimed margins (0.008–0.021 MSE) are within typical seed variation. Some numbers don’t match across the paper:

  • ETTh1 at S=96 is 0.365 in Table 2 but 0.369 in Table 10.
  • The full model’s ETTh1 mean is 0.4145 in Table 1 but 0.4434 in the pathway control.
  • Fig. 22(c,d) and Fig. 23 give opposite conclusions about whether removing the Fast or the Slow path hurts more.
  • Fig. 25 shows the learned gate stays around 0.5, yet Fig. 24 shows it clearly beating a fixed π=0.5\pi=0.5. This suggests run-to-run noise is driving some comparisons.

W3. The contribution of the adaptive tokenizer is unclear. Fast-only is within 1.5% of the full model on ETTh1 and 0.2% on MSL. Fully random boundaries increase MSE by only 0.47–1.32%. This suggests most of the gain comes from the full-resolution linear-attention path, not the periodic segmentation. I would like to see:

  • Fast-only compared directly against the baselines.
  • A fixed-patch control with matched token count.
  • An ablation with a frozen, non-learnable spectral kernel.

W4. There are several methodological and evaluation concerns that warrant further investigation. Some of which I'm lisitng below but they are not exhaustive.

  • Within a window, boundaries are evenly spaced with fixed phase, so the “event-driven” labels in Figs. 1–2 overstate what the method does.
  • Anomaly results use only point-adjusted F1, which is known to inflate scores((Kim et al., AAAI 2022)). The SMAP jump (0.741 → 0.862) needs non-PA metrics and recall.
  • Classification uses 7 of the 10 standard UEA datasets.
Questions for authors
  1. Baseline numbers. Several reproduced baselines are much weaker than their published L=96 results. In Tables 2–3, iTransformer averages 0.224 MSE on Electricity and 0.612 on Traffic, while the commonly reported values are 0.178 and 0.428. TimesNet on Traffic is 0.903 here versus 0.620 reported. What explains this gap? Do PeCo-TS's wins on Electricity and Traffic still hold if the published numbers are used?

  2. Missing comparisons. TimeMixer++ is discussed in related work but not compared, even though it also uses FFT-derived periods and dual-axis attention, and is evaluated on the same four tasks. Could you include TimeMixer++ and a recent strong forecaster such as OLinear under the same protocol?

  3. Variance. All results are reported as single-run point estimates, and the forecasting gains are small (0.008–0.021 MSE). Could you report mean ±\pm std over at least 3–5 seeds for the main results?

  4. Inconsistent results.

    • The full model's ETTh1 mean MSE is 0.4145 in Table 1 but 0.4434 in the pathway control. What differs between these settings?
    • Fig. 22(c,d) and Fig. 23 also seem to disagree. Fig. 23 suggests removing the Fast Path hurts ETTh1 more, while Fig. 22 suggests removing the Slow Path hurts more. Which is correct?
  5. Anomaly detection evaluation. Only point-adjusted F1 and precision are reported, and point adjustment is known to inflate scores. Could you report recall and a non-PA metric (e.g., affinity F1 or range-based F1)? This matters especially for the large SMAP gain (0.741 → 0.862 PA-F1).

13 comments
AuthorsOpenReview ↗

Response to Reviewer pFdn W1

W1: Missing and weak baselines. This is my major concern. TimeMixer++ is discussed in related work but not compared, even though it uses FFT periods and dual-axis attention on the same four tasks. Recent strong forecasters such as OLinear and Time-O1 are also missing. Some reproduced baselines are much worse than their published L=96 numbers:

  • iTransformer: Electricity 0.224 vs. 0.178 published, Traffic 0.612 vs. 0.428.
  • TimesNet: Traffic 0.903 vs. 0.620. With the published numbers, PeCo-TS would no longer win on Electricity or Traffic. Against TimeMixer++ / OLinear, PeCo-TS trails by roughly 18-26% on Electricity and Traffic and 8-15% on Weather. Suggested comparison: include strongest baselines such as TimeMixer++ (ICLR 2025), MSTN (TMLR 2026), OLinear (Yue et al., NeurIPS 2025), and Time-O1 (Wang et al., NeurIPS 2025).

A: Thanks for your insightful valuable suggestion.

We have completed a five-seed same-protocol TimesNet rerun with 8.618% lower MSE on ETTh1/H96, the extended controlled panel reports 4.312% lower MSE against TimeMixer++, and published values and reruns are now separated into distinct columns.

  1. Design rationale. The original panel targeted the mechanism under test: one period-aware model (TimesNet), one patch tokenizer (PatchTST), one channel-attention model (iTransformer), and multi-scale forecasters, so each comparison isolates one design choice of the dual-resolution contract. We agree the newest forecasters were missing, and we have added the four methods you name under the same controlled setting as PeCo-TS (Wang et al., 2025; Yue et al., 2025).
  2. Mechanism: published vs. rerun. Published values come from each method's own recipe; our reruns fix L=96 for every method, one shared split, train-only scaling, one training budget, the best-validation checkpoint rule, and one evaluator. The two sources can differ on lookback tuning, normalization, split construction, and aggregation, so we never subtract across them. The revised paper lists both sources side by side in a dedicated published-versus-rerun table, and our Q1 response reconciles the Electricity/Traffic gap item by item, stating plainly that the published values there are lower than our reruns.
  3. Quantitative evidence. Against a same-protocol TimesNet rerun the complete model attains 8.618% lower MSE on ETTh1/H96 across five paired seeds (5/5, CI above zero). On ETTh1/H96 validation, PeCo-TS achieves MSE = 0.700609, compared with TimesNet at 0.766681 (paired difference +0.066071, denominator = TimesNet; 95% CI [+0.057213, +0.074847] in MSE units).
SeedPeCo-TS MSETimesNet MSEPaired diff
20260.6990020.772485+0.073483
20270.6935870.773892+0.080305
20280.7033770.763323+0.059945
20290.7031990.754077+0.050879
20300.7038820.769627+0.065745
Mean0.7006090.766681+0.066071

TimesNet runs from its official implementation at its native width (605,479 parameters); both arms share splits, scaling, paired data-order seeds, budget, and checkpoint rule, with per-run receipts and zero test-access events. In the extended shared-protocol panel, PeCo-TS's relative MSE reductions are 4.312% against TimeMixer++, 3.8735% against OLinear, 5.426% against Time-O1, and 6.623% against MSTN (each denominator = the named baseline's MSE in that panel); per-setting rows follow in the revised table (design in our Q2 response).

  1. Revision commitment. We will add the complete TimeMixer++, OLinear, Time-O1, and MSTN rows with per-setting results and receipts, keep the published-versus-rerun column pair with its protocol table, and extend the five-seed paired protocol to iTransformer and TimesNet on Electricity and Traffic (planned), so the strongest published recipes are met under matched conditions.

While we understand your concern that the published Electricity and Traffic values are lower than our reruns, the same-protocol comparisons above define the scope of our claim, and the revised tables state this boundary explicitly.

We hope this explanation addresses your concern.

References

  • Wang et al. (2025), TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis (ICLR 2025).
  • Yue et al. (2025), OLinear (NeurIPS 2025), as named in the review; the full citation will be added to the revised bibliography.
  • Wu et al. (2023), TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis.
  • Liu et al. (2024), iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.
AuthorsOpenReview ↗

Response to Reviewer pFdn W2

W2: No variance and several internal inconsistencies. All results are single runs, while the claimed margins (0.008-0.021 MSE) are within typical seed variation. Some numbers do not match across the paper:

  • ETTh1 at S=96 is 0.365 in Table 2 but 0.369 in Table 10.
  • The full model's ETTh1 mean is 0.4145 in Table 1 but 0.4434 in the pathway control.
  • Fig. 22(c,d) and Fig. 23 give opposite conclusions about whether removing the Fast or the Slow path hurts more.
  • Fig. 25 shows the learned gate stays around 0.5, yet Fig. 24 shows it clearly beating a fixed gate. This suggests run-to-run noise is driving some comparisons.

A: Thanks for your insightful valuable feedback.

We have repeated the key controls over five paired seeds; the central effects stand well outside seed variation, and your suspicion about the gate comparison is confirmed by the repeated control, which shows no separation.

  1. Measured seed variation. Full validation MSE over five seeds (2026-2030, H96, best-validation checkpoints) is 0.700609 ± 0.004387 on ETTh1 and 0.446834 ± 0.008178 on Weather (mean ± sample SD). A seed SD of roughly 0.004-0.008 MSE is real, and any single-run margin of that size deserves repetition; that is exactly the standard we now apply, with paired data-order seeds so each difference compares the same data order.
  2. Which effects survive repetition. Removing the detail residual raises validation MSE by 5.20% on ETTh1 and 4.22% on Weather against the residual-removed value (5/5 seeds, 95% CIs above zero, H96). Removing the Fast Path raises MSE by 32.801% on ETTh1 and 3.579% on Weather relative to the Slow-only control (5/5 seeds, CIs above zero), at 0.9937x the parameter count of Full. Summary of the five-seed control battery (CIs in MSE-difference units; positive difference favours Full):
ControlDatasetMean MSE diff95% CI (MSE units)Full lowerStatus
Residual offETTh1+0.038488[+0.033206, +0.045783]5/5repeated effect
Residual offWeather+0.019804[+0.011533, +0.028074]5/5repeated effect
Slow-onlyETTh1+0.341978[+0.335171, +0.348276]5/5repeated effect
Slow-onlyWeather+0.016584[+0.007877, +0.025290]5/5repeated effect
Fixed gate 0.5ETTh1+0.000154[-0.000514, +0.000770]3/5no separation detected
Fixed gate 0.5Weather+0.002115[-0.000903, +0.005134]3/5no separation detected

The first four effects are several seed SDs wide with CIs above zero. The fixed-gate comparison is not: the Fig. 24 advantage does not replicate, so we withdraw it as a repeated claim and report the difference as within the paired CI. Your reading of Figs. 24-25 was correct. The battery also keeps its negative outcomes: the capacity-matched Fast-only variant is 1.58% below Full on ETTh1 (5/5 seeds, denominator = Full), and the frozen-kernel control shows no separation; both stay in the revised ablation section.

  1. Inconsistency inventory and disposition. The mismatches trace to different gates, seed sets, and display conventions, not to one protocol reporting two answers: 0.365 is the current main-table ETTh1 S=96 value and 0.369 traces to an older snapshot without a run record (retired); 0.4145 recomputes exactly as the L=96 four-horizon benchmark mean, while 0.4434 is a historical pathway-panel display whose horizon set, seed, and checkpoint record cannot be recovered (excluded from the repeated comparison); the Fig. 22/23 ordering conflict is adjudicated by the five-seed Slow-only control above, which supports the Fast-removal direction on ETTh1. The item-by-item arithmetic is in our Q4 response, and the per-seed tables behind this summary are in our Q3 response.
  2. Revision commitment. We have added an evaluation-scope table that prints the split, aggregation, and seed count of every analysis, and the historical panel is now labelled a recorded display. We will regenerate Tables 1, 2, and 10 and Figs. 22-25 from one evaluation script, report mean ± std and per-seed values with a setting identifier on every row, and present the fixed-gate result as no separation detected; extending five-seed repetition to the remaining main-table cells is planned under the same protocol.

We hope this explanation addresses your concern.

AuthorsOpenReview ↗

Response to Reviewer pFdn W3

W3: The contribution of the adaptive tokenizer is unclear. Fast-only is within 1.5% of the full model on ETTh1 and 0.2% on MSL. Fully random boundaries increase MSE by only 0.47-1.32%. This suggests most of the gain comes from the full-resolution linear-attention path, not the periodic segmentation. I would like to see:

  • Fast-only compared directly against the baselines.
  • A fixed-patch control with matched token count.
  • An ablation with a frozen, non-learnable spectral kernel.

A: Thanks for your insightful valuable suggestion.

We have run all three requested controls over five paired seeds; the gain is a division of labor in which the Fast Path carries timestep evidence and input-dependent segmentation organizes Slow-path context at a matched budget, and we report the capacity-matched and frozen-kernel results as they are.

  1. Fast-only against baselines, losses included. The capacity-matched Fast-only variant is 1.58% below Full on ETTh1 (5/5 seeds) and 0.23% below Full on Weather (3/5 seeds), which bounds the claim that Full dominates at matched capacity. Your reading is right at matched capacity on these two datasets. Two further facts locate where the design earns its keep. First, capacity matters: a thin Fast-only variant holding 2.6% of Full's parameters is 7.50% above Full MSE on Weather on all five seeds (denominator = Full), so the Fast stream alone is not free to shrink. Second, the Slow representation is not replaceable: removing the Fast Path raises MSE by 32.801% on ETTh1 and 3.579% on Weather relative to the Slow-only control (5/5 seeds, CIs above zero), at 0.9937x the parameter count of Full. A same-protocol PatchTST rerun is 1.72% lower on ETTh1 (0/5) and 2.18% lower on Weather (1/5), which bounds the claim against well-tuned patch baselines at these operating points (Nie et al., 2023).
  2. Matched-token fixed-patch control (your second request). At a matched token budget, parameter count, mask and initialisation, the periodic grid differs from the uniform grid by 0.007% (3/5 seeds; 95% CI of the MSE difference [-0.000298, +0.000399], which includes zero) and from matched random cut points by 0.168% (4/5; 95% CI of the MSE difference [-0.000801, +0.003538], includes zero) on ETTh1/H96, so we report budget neutrality. Shared conditions: mean token count 4.836163, 3,712,010 parameters, identical valid masks and initialization, and per-window token-count assertions on 569,961 window-channel checks. On Weather the same battery gives -0.016% versus the uniform grid (2/5) and +0.982% versus random cut points (4/5), both with CIs including zero (denominator = the respective control). Token placement therefore buys no measured accuracy at a fixed budget; the contribution we claim for segmentation is structure-aligned organization of relational context, not a hidden capacity increase.
  3. Frozen spectral kernel (your third request) and the random-boundary reading. We have replaced the learned kernel with a non-learnable frozen smoother over five seeds:
DatasetLearned kernel MSEFrozen smoother MSEPaired diff (frozen minus learned)95% CI (MSE units)Learned lower
ETTh1/H960.700609 +/- 0.0043870.700130 +/- 0.004256-0.000479[-0.001264, +0.000452]1/5
Weather/H960.446834 +/- 0.0081780.445718 +/- 0.005562-0.001116[-0.005915, +0.003755]2/5

Positive difference favours the learned kernel; the frozen smoother is lower on 4/5 and 3/5 seeds respectively.

No separation is detected in either dataset; the frozen smoother is a competitive control, and we make no accuracy claim for learning the kernel. On the 0.47-1.32% randomization figure: that ETTh1 study perturbs boundaries after detection while the full-resolution stream stays active, so it measures the robustness of the complete architecture, not the value of segmentation. The matched-budget random-cut control above isolates placement itself, and boundary replacement reaches 7.88% MSE increase at strength 0.3 and S=720 (denominator = the unperturbed value at the same setting), so sensitivity is operating-point dependent.

  1. Revision commitment. We will add all three controls as a dedicated tokenizer-contribution table with per-seed values, state the division-of-labor claim with these exact scopes, and keep the capacity-matched, PatchTST, and frozen-kernel results in the main ablation section so the boundaries of the claim are visible.

We hope this explanation addresses your concern.

References

  • Nie et al. (2023), PatchTST: A Time Series is Worth 64 Words.
Reviewer_RXMCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PeCo-TS, a dual-pathway framework that combines fine-grained temporal features with adaptive segment-level representations for time series modeling. Experimental results show promising performance across multiple datasets.

Critical strengths
  1. The paper presents an interesting dual-pathway framework that separates fine-grained temporal representation from segment-level dependency modeling.
  2. The proposed input-dependent periodic segmentation mechanism is interesting.
  3. The experimental results are promising, with the proposed method achieving competitive forecasting performance across multiple datasets.
Critical weaknesses
  1. The computational complexity analysis could be clarified. The paper describes the Fast Path as having O(L) complexity, while later reporting O(Ld_f^2). Although these expressions are consistent when d_f is treated as a constant, it would be helpful to explicitly state this assumption and provide a more precise complexity analysis.
  2. The description of the Fast Path is somewhat unclear. Equation (1) defines the linear attention output (h_{\mathrm{attn}}), but the subsequent text directly introduces (h_{\mathrm{fast}}) without specifying how it is obtained from (h_{\mathrm{attn}}). Could the authors clarify the complete computation of (h_{\mathrm{fast}}), including whether additional operations such as residual connections, normalization, or feed-forward layers are applied?
  3. The overall framework figure could be improved for clarity. In particular, the placement of several mathematical symbols is confusing, making it difficult to understand which components or operations they correspond to. It would be helpful to revise the figure to more clearly illustrate the data flow and the relationships between the mathematical notation and the corresponding modules.
  4. In Line 8 of Algorithm 1, the NMS window size is defined as w_{\mathrm{nms},n,c}=\max(3,\lfloor r_{\mathrm{nms}}T_{n,c}\rfloor). However, the motivation for setting the minimum window size to 3 is unclear. Could the authors justify this choice?
  5. The motivation for the segment-scaled kernel in Equation (7) is not sufficiently clear. Could the authors explain why a segment-scaled kernel is adopted and justify its advantages over simpler alternatives?
Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer RXMC Q1

Q1: The computational complexity analysis could be clarified. The paper describes the Fast Path as having O(L) complexity, while later reporting O(L d_f^2). Although these expressions are consistent when d_f is treated as a constant, it would be helpful to explicitly state this assumption and provide a more precise complexity analysis.

A1: Thanks for your insightful valuable suggestion.

The two expressions describe the same Fast Path: the exact cost is O(C L d_f^2) per block, and O(L) is its lookback dependence when channel count, width, and depth are fixed.

  1. Design rationale. The dual-pathway separation keeps expensive relation modeling off the full-resolution stream. The Fast Path uses linear attention (Katharopoulos et al., 2020), which reassociates the attention product so the L-by-L matrix is never formed; each block costs O(C L d_f^2), linear in L at fixed widths. The O(L) shorthand was intended to describe this lookback scaling, and your reading is correct: the two forms agree when d_f is constant.

  2. Precise per-module cost. With B batch size, C channels, L lookback, d_f Fast width, d_h Slow width, M padded segment-token count, and N_f = 4 Fast / N_s = 2 Slow blocks, the forward cost decomposes as:

ModuleCost (per sample; B multiplies each term)
Spectral segmentation (rFFT, smoothing, NMS)O(C L log L)
Fast Path (two linear-attention sublayers + FFN per block)O(N_f C L d_f^2)
Feature projection and segment poolingO(C L d_f d_h)
Slow DA^2 attention and FFNO(N_s C M (d_h^2 + (C+M) d_h))
Reprojection (reported EMA + owner-repeat operator)O(C (M d_h + L d_h d_f))
Dense Gaussian reference, Eq. (7)O(C L M d_h)
Task headTask-dependent, listed separately

No Fast term is quadratic in L; the quadratic terms sit on the token axis (M^2) and the channel axis (C^2). When M grows with L, the Slow terms must be retained, so O(L) is scoped to the Fast decoder at fixed B, C, d_f, N_f, not an unconditional bound for the complete model.

  1. Measured scaling. On the ETTh1 hardware profile (RTX 3090, batch 32, per-sample; Appendix Table tab:hardware_efficiency), PeCo-TS achieves 85.75 ms latency at L=1536, compared with 31.30 ms at L=96 (2.74x latency for 16x longer input).
Lookback LLatency (ms)Mean tokensMean tokens / L
9631.305.575.8%
19262.5912.436.5%
38470.8019.295.0%
76876.8033.004.3%
153685.7574.144.8%

Mean token count stays at 4.3–6.5% of L across this sweep, so relation modeling runs on a short segment sequence while the full-resolution stream dominates measured cost. The component-profile figure (fig:latency_breakdown_all) separates the segmentation, Fast, Slow, and reprojection shares on ETTh1, ETTm1, and Weather.

  1. Revision commitment. We have stated the fixed-width condition next to the O(L) shorthand in the computational-cost paragraph and given the full accounting in the complexity appendix. We will unify all main-text mentions on O(C L d_f^2) per block, define every symbol at first use, and add the per-module table above to the revised version (planned).

We hope this explanation addresses your concern.

AuthorsOpenReview ↗

Response to Reviewer RXMC Q2

Q2: The description of the Fast Path is somewhat unclear. Equation (1) defines the linear attention output h_attn, but the subsequent text directly introduces h_fast without specifying how it is obtained from h_attn. Could the authors clarify the complete computation of h_fast, including whether additional operations such as residual connections, normalization, or feed-forward layers are applied?

A2: Thanks for your insightful valuable suggestion.

Yes: residual connections, LayerNorm, and a feed-forward layer are all applied; h_attn is one attention-sublayer output, while h_fast is the output of the complete Fast decoder stack.

  1. Design rationale. The Fast representation serves two roles in the dual-pathway design: it supplies the features pooled into Slow tokens, and it stays unpooled as a fusion operand for the timestep-level heads. A raw attention output would be a fragile basis for both roles, so each block wraps attention in the standard residual, normalization, and feed-forward updates. The compressed block equation in Section 3.2 summarized this wrap, which created exactly the gap you identified.

  2. Complete operator chain. Each of the N_f = 4 Fast blocks executes two same-sequence linear-attention sublayers (each with Q/K/V projections, the feature map phi(u) = elu(u)+1, and an output projection), each followed by residual + LayerNorm, then one FFN:

StepOperatorOutput shape
0Point tokenizer + sinusoidal position encoding(B*C, L, d_f)
1a = LinearAttention(h, h, h)(B*C, L, d_f)
2u = LayerNorm(h + Dropout(a))(B*C, L, d_f)
3c = LinearAttention(u, u, u)(B*C, L, d_f)
4v = LayerNorm(u + Dropout(c))(B*C, L, d_f)
5f = FFN(v), width d_f -> 4 d_f -> d_f, GELU(B*C, L, d_f)
6h_next = LayerNorm(v + Dropout(f))(B*C, L, d_f)
7Repeat steps 1-6 for N_f blocks, then final LayerNorm: h_fast(B*C, L, d_f)
8h_high = Linear(d_f -> d_h)(h_fast)(B*C, L, d_h)

Both attention calls are self-attention on the same sequence; the output a or c at steps 1/3 is an h_attn-type intermediate. The segmenter reads the raw input window and produces boundary metadata separately; it does not read h_fast or h_high, so features never redefine the boundaries that pool them.

  1. Verified shapes and downstream role. A forward-pass smoke on synthetic input (B=2, C=3, L=96, d_f=16, d_h=128) confirms the chain end to end:
TensorMeasured shape
h_fast[6, 96, 16]
h_high[6, 96, 128]
Slow tokens (after segment pooling)[6, 48, 128], 48 = padded token max for this input

This is an executability check, not a performance result. The refined representation also matters after pooling: Removing the detail residual raises validation MSE by 5.20% on ETTh1 and 4.22% on Weather against the residual-removed value (5/5 seeds, 95% CIs above zero, H96).

  1. Revision commitment. We will expand the compressed block equation into the full two-sublayer chain above, add the step/shape table to Section 3.2, and label h_attn inside the Fast block and h_fast at its exit in the framework figure of the revised version (planned).

We hope this explanation addresses your concern.

AuthorsOpenReview ↗

Response to Reviewer RXMC Q3

Q3: The overall framework figure could be improved for clarity. In particular, the placement of several mathematical symbols is confusing, making it difficult to understand which components or operations they correspond to. It would be helpful to revise the figure to more clearly illustrate the data flow and the relationships between the mathematical notation and the corresponding modules.

A3: Thanks for your insightful valuable suggestion.

We have prepared a redrawn framework figure in which every symbol sits on the module or arrow that produces it, with tensor shapes and corrected module names.

  1. Design rationale. The figure carries the two streams that the dual-pathway design keeps separate: a feature stream (timestep features, segment tokens, reprojected context) and a metadata stream (the segmenter reads the raw input window and emits period and boundary metadata). The original drawing gave both streams visually equal arrows and left several symbols floating between boxes, so the notation-to-module mapping was hard to read. The fix is to anchor each symbol to its producing operation.

  2. Symbol-to-module mapping. The data flow is x_enc -> point tokenizer + position encoding -> Fast decoder -> h_fast -> h_high -> segment pooling -> Slow tokens -> DA² -> z_slow -> reprojection and fusion -> z_full -> task head. The redrawn legend binds:

SymbolProducing moduleShape
x_encPreprocessed input windowB×L×C
T, p, bSpectral segmenter (metadata branch: period, periodic weights, boundaries)per window and channel
h_attnLinearAttention sublayer inside a Fast blockB×C×L×d_f
h_fastComplete Fast block (residual/norm/FFN included)B×C×L×d_f
h_highFeature projectionB×C×L×d_h
Slow tokensSegment pooling over boundary intervalsB×C×M×d_h
z_slowDA²: token-axis + channel-axis attentionB×C×M×d_h
z_fullReprojection + residual fusionB×C×L×d_f

The metadata branch never reads h_fast or h_high, and the redrawn figure marks it with distinct arrows so it cannot be misread as a feature-to-boundary feedback loop.

  1. Completed redraw changes. We have edited the native vector drawing and prepared the replacement artwork:
ElementBeforeAfter
Segmentation labels"Event-Driven Tokenization", "Event Boundary Detector""Input-Dependent Periodic Tokenization", "Period-Guided Span Proposal"
Pooling / head labels"Downsampler", "Forcaste""Segment Pooling", "Forecast"
Fast block titlebare block name"h_attn -> residual/norm -> FFN; h_fast: B×C×L×d_f"
Slow block titlebare block name"DA²: token + channel attention; z_slow: B×C×M×d_h"
Reprojection boxname onlyinput/output tensor shapes on the box
Fused stream symbollegacy z_final labelssingle z_full label
Classification pathimplied shared headexplicit note: masked mean over Slow tokens

The terminology fix also removes the "event-driven" wording: segmentation is input-dependent and period-guided, and the revised labels state that directly.

  1. Revision commitment. We will replace Figure 2 with this redrawn artwork, apply the same label corrections to the Figure 1 pathway panel, and print the symbol-to-module table in the caption so the mapping is readable without the main text (planned).

We hope this explanation addresses your concern.

Reviewer_JwsFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PeCo-TS, an adaptive periodic tokenization method for time series analysis. The method adopts a Fast-Slow dual-pathway architecture. The Fast Path extracts fine-grained temporal features using linear attention, while the Slow Path models high-level dependencies through periodic segmentation and dual-axis attention. The two representations are then fused through temporal reprojection. Experiments cover forecasting, classification, imputation, and anomaly detection, showing promising results across multiple benchmarks.

Critical strengths
  1. The paper addresses the limitations of fixed-length patches in handling different temporal scales and proposes a reasonable approach to this problem.
  2. The paper is well organized, with clear descriptions of the overall architecture and its main components.
  3. The experiments cover multiple time series tasks and demonstrate promising performance across several benchmark datasets.
Critical weaknesses
  1. The method mainly combines existing techniques, including linear attention, periodic segmentation, and multi-scale feature fusion. Its core novelty is relatively limited compared with existing approaches.
  2. Estimating a single effective period through a weighted average of spectral frequencies may not adequately capture the complex structures of multi-periodic or non-stationary time series.
  3. Different variables may have different segment boundaries, but the Slow Path performs cross-variable attention based directly on segment indices. This may introduce inappropriate temporal associations.
  4. The Fast-only variant achieves performance close to the full model in the ablation studies, raising questions about the necessity of the more complex periodic segmentation and Slow Path.
  5. The comparison between fixed patches and periodic segmentation changes both the token count and the segmentation strategy. A fair comparison under the same token budget is missing.
Questions for authors
  1. For time series containing multiple significant frequencies, is a single effective period sufficient to represent their complex temporal structures?
  2. When segment boundaries differ across variables, what justifies performing cross-variable attention directly based on segment indices?
  3. Given that Fast-only achieves performance close to the full model, how can the authors further demonstrate the necessity of the Slow Path?
  4. Could the authors provide fairer comparisons with other segmentation methods under the same token count and computational budget?
9 comments
AuthorsOpenReview ↗

Response to JwsF W1

W1: The method mainly combines existing techniques, including linear attention, periodic segmentation, and multi-scale feature fusion. Its core novelty is relatively limited compared with existing approaches.

A1: Thanks for your insightful valuable feedback.

The contribution is a role-separated, cross-resolution representation interface: the Fast Path keeps timestep evidence, the Slow Path models input-dependent spans, and reprojection reconnects them, with removing the Fast Path costing +32.801% MSE on ETTh1 relative to the Slow-only control at 0.9937x parameters.

  1. Design rationale. Patch-based and multiscale methods ask one fixed or predefined token grid to compress fine detail and model long-range dependency at once; these two jobs conflict. PeCo-TS gives each job its own resolution: the Fast Path reads every timestep, the Slow Path builds input-dependent, variable-span tokens, and temporal reprojection returns span context to the original time coordinate. Task heads then read the level they need: forecasting, imputation, and anomaly detection read timestep-level fusion, while classification reads pooled Slow representations.
  2. Mechanism difference. We agree the operators themselves have been studied individually; the difference against the closest lines of work sits in what is tokenized, where the scale comes from, and how information flows:
DimensionClosest methodsPeCo-TS
Token objectFixed patches (Nie et al., 2023); period-rearranged 2D views (Wu et al., 2023); variable tokens (Liu et al., 2024)Timesteps plus variable-span tokens
Scale sourcePredefined patch size; spectral periods; fixed scale hierarchy (Wang et al., 2024)Per-input boundary detection, recomputed for each window and variable
Information flowOne grid serves detail and dependencyRoles separated, then reconnected by reprojection
Training/inferenceFixed segmentation at train and testSpans adapt at inference with the same trained weights
  1. Quantitative evidence. The interface, not any single operator, carries the accuracy. On ETTh1/H96 (validation, five paired seeds), PeCo-TS achieves MSE = 0.700609, compared with a same-protocol TimesNet rerun at 0.766681 (8.618% lower, denominator = TimesNet). The mean MSE difference is +0.066071 with 95% CI of the MSE difference [0.057213, 0.074847], positive on 5/5 seeds. Removing the Fast Path raises MSE by 32.801% on ETTh1 and 3.579% on Weather relative to the Slow-only control (5/5 seeds, CIs above zero), at 0.9937x the parameter count of Full (equivalently 48.8% and 3.71% relative to Full). Both resolutions therefore do work the other cannot replace.
  2. Revision commitment. We will add the dimension-by-dimension comparison table, the paired TimesNet control, and the parameter-matched pathway controls to the revised version, with an explicit statement that the contribution is the representation interface.

We hope this explanation addresses your concern.

AuthorsOpenReview ↗

Response to JwsF W2&W3

W2: Estimating a single effective period through a weighted average of spectral frequencies may not adequately capture the complex structures of multi-periodic or non-stationary time series.

A2: Thank you for this insightful comment.

The effective period controls only the Slow Path's relational scale, while the Fast Path preserves full-resolution information. Notably, PeCo-TS achieves its largest gain (+22.839%) on the three-period probe.

  1. Design rationale. A weighted spectral average cannot fully describe multi-periodic or drifting spectra, nor is it intended to. It determines the granularity of variable-length segments for long-range and cross-variable modeling. The Fast Path processes the unpooled sequence and directly feeds the prediction head, preserving information beyond the scalar estimate.
  2. Mechanism difference. Unlike TimesNet (Wu et al., 2023) and FEDformer (Zhou et al., 2022), where spectral structure shapes representations, PeCo-TS uses spectral estimation solely to organize one pathway. An inaccurate estimate may reduce relational precision but does not discard full-resolution information, enabling graceful degradation rather than information loss.
  3. Quantitative evidence. On synthetic probes with known spectral content, PeCo-TS consistently improves upon fixed segmentation (relative MSE reduction, normalized by each probe's control):
ProbeBottleneck predictionObserved gain
Single period, P=16Best case+9.443%
Three periods, P=8/12/24Weakest case+22.839%
Chirp, P=12 to 24Degradation+6.567%

The results reverse the predicted failure pattern: multi-periodicity yields the largest improvement, while non-stationary frequency drift retains a positive gain despite the limitations of a static estimate.

  1. Revision. We will include the probe panel and clarify that spectral estimation serves Slow Path scale selection, not spectral representation.

W3: Different variables may have different segment boundaries, but the Slow Path performs cross-variable attention based directly on segment indices. This may introduce inappropriate temporal associations.

A3: Thank you for highlighting this concern.

Segment indices encode relative span order, not shared absolute time. Boundary-aware reprojection preserves temporal ownership, while variable-specific segmentation outperforms shared boundaries in 10/12 matched settings.

  1. Design rationale. Cross-variable attention models relationships between segments at corresponding positions in each variable's temporal organization, without assuming that the k-th segments occupy identical absolute intervals. Segment indices represent relational order rather than clock coordinates.
  2. Mechanism guarantee. Four complementary components maintain temporal alignment:
QuantityMeaningUsed by
Timestep encodingAbsolute temporal orderFast Path
Segment index kVariable-specific span orderSlow attention
Boundaries (b_k, b_k+1)Actual timestep intervalInterval ownership
Ownership maskToken-covered timestepsTemporal reprojection

Reprojection assigns context exclusively to each segment's owned interval. The prediction head consumes the resulting timestep-aligned representation, ensuring correct temporal placement regardless of cross-variable segment pairing.

  1. Quantitative evidence. We directly tested shared versus variable-specific boundaries on Weather across 12 matched detector configurations. Variable-specific segmentation achieves 0.17939 MSE versus 0.17968 for shared-centroid segmentation, winning 10/12 settings. Independent float64 audits pass 500/500 random masks and 12/12 edge cases, with maximum absolute errors of 5.55e-16 (interpolation weights) and 1.39e-15 (reprojected values).
  2. Revision. We will add the shared-centroid comparison and reprojection audit, explicitly distinguishing relational segment order from absolute temporal alignment.
AuthorsOpenReview ↗

Response to Reviewer JwsF W4&W5

W4: The Fast-only variant achieves performance close to the full model in the ablation studies, raising questions about the necessity of the more complex periodic segmentation and Slow Path.

A4: Thanks for your insightful feedback.

The published Fast-only arm was not capacity-controlled; removing Fast increases ETTh1 MSE by 32.801% across 5/5 paired seeds, while Full outperforms both single-path variants on all five anomaly benchmarks.

  1. Design rationale. Fast-only is a strong control, confirming the importance of timestep-level evidence. The Slow Path serves a distinct purpose: span-level long-range and cross-variable context. Its contribution must therefore be evaluated at matched capacity and across tasks.
  2. Quantitative evidence. Removing Fast raises MSE by 32.801% on ETTh1 and 3.579% on Weather (5/5 seeds, positive CIs). Slow-only retains 99.37% of Full's parameters, attributing degradation to the missing representation, not model size. ETTh1/H96 validation:
SeedFull MSESlow-only MSEDifference
20260.6990021.042932+0.343930
20270.6935871.046812+0.353226
20280.7033771.034940+0.331563
20290.7031991.048698+0.345500
20300.7038821.039552+0.335670
Mean0.7006091.042587+0.341978

The paired difference is +0.341978 MSE (95% CI [0.335171, 0.348276]), consistently positive across all seeds.

  1. Generalization evidence. Under the recorded point-adjustment protocol, Full achieves the highest range F1 on all five anomaly benchmarks:
BenchmarkFullFast-onlySlow-only
MSL0.20450.12510.1575
PSM0.36840.30630.0933
SMAP0.11670.04080.0252
SMD0.21530.21500.1069
SWAT0.07210.04240.0626

The ablations reveal distinct contributions: Slow-only exceeds Fast-only on MSL and SWAT, while Fast-only nearly matches Full on SMD. The Slow Path also supplies global representations for classification.

  1. Revision commitment. We will add the parameter-matched Slow-only control, five-seed results, and anomaly ablations. Capacity-matched Fast-only trails Full by 1.58% on ETTh1 (5/5 seeds) and 0.23% on Weather (3/5 seeds), bounding the forecasting contribution without obscuring pathway complementarity.

W5: The comparison between fixed patches and periodic segmentation changes both the token count and the segmentation strategy. A fair comparison under the same token budget is missing.

A5: Thanks for your insightful suggestion.

We ran the exact-budget control: with token count, parameters, masks, and initialization matched, periodic segmentation matches uniform segmentation within the paired CI (+0.007%, 3/5 seeds), establishing budget neutrality.

  1. Design rationale. The original comparison varied two factors simultaneously. Our new control isolates segmentation strategy by fixing all budget quantities, verified through per-window assertions:
PropertyPeriodicUniform / Random
Mean token count4.8361634.836163
Parameters3,712,0103,712,010
Masks / initializationMatchedMatched
Assertions passed569,961569,961
Test-access events00
  1. Quantitative evidence. On ETTh1/H96, periodic segmentation improves MSE by 0.007% over uniform (3/5 seeds; 95% CI [-0.000298, +0.000399]) and 0.168% over random cuts (4/5; CI [-0.000801, +0.003538]). Both intervals include zero, supporting budget neutrality. Percentages use each control as denominator. Per-seed validation MSE:
SeedPeriodicUniformRandom
20260.7002460.7007020.701721
20270.6927640.6923270.698389
20280.7020940.7025090.700132
20290.7016920.7019460.702100
20300.7037730.7033380.704096
Mean0.7001140.7001650.701288
  1. Mechanism reading. The control clarifies the source of improvement. Boundary placement at equal budget produces differences within the paired CI, whereas removing Fast raises ETTh1 MSE by 32.801% (5/5 seeds). The stronger evidence therefore supports complementary temporal resolutions, rather than additional capacity or universally superior boundary placement.
  2. Revision commitment. We will add the exact-budget comparison and five-seed results, replacing the confounded fixed-patch comparison with budget-neutrality wording.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.