Pith. sign in

REVIEW 2 major objections 5 minor 50 references

Knowledge distillation helps intermediate-size time-series models most, cutting parameters sharply while keeping accuracy competitive.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 21:09 UTC pith:ZLXHHZNA

load-bearing objection Solid large-scale empirical extension of response-based KD to three TSC architectures; intermediate-complexity students win consistently, with public code and no load-bearing flaw. the 2 major comments →

arxiv 2607.06796 v1 pith:ZLXHHZNA submitted 2026-07-07 cs.LG cs.AI

Enhancing deep learning models for time series classification via knowledge distillation

Javidan Abdullayev , Maxime Devanne , Jonathan Weber , Germain Forestier This is my paper
classification cs.LG cs.AI
keywords time series classificationknowledge distillationmodel compressionFCNInceptionTimeConvTranUCR Archive
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Deep time-series classifiers often become too large for phones, sensors, or other constrained devices. This paper tests whether knowledge distillation—training a smaller student to mimic a larger teacher’s soft predictions—can shrink three popular architectures (FCN, Inception, and ConvTran) without a large accuracy drop. On 112 UCR datasets the pattern is consistent: students of intermediate complexity gain the most from the teacher, while very large students already match the teacher on their own and very small ones often cannot absorb the knowledge. Concrete wins include an FCN student with 38× fewer parameters, an Inception student that nearly matches its teacher with 42% fewer parameters, and a two-head ConvTran student that improves most over its undistrilled twin. The work also shows that distilled convolutional filters become more like the teacher’s and that distillation reduces overfitting, acting as a regularizer.

Core claim

Across FCN, Inception and ConvTran architectures evaluated on 112 UCR time-series datasets, knowledge distillation most effectively improves student models of intermediate complexity; the best FCN student cuts parameters by a factor of 38, the best Inception student nearly matches the teacher with 42% fewer parameters, and the two-head ConvTran student shows the largest gain over its undistrilled counterpart.

What carries the argument

Response-based knowledge distillation: the student is trained on a weighted sum of ordinary cross-entropy against hard labels and KL divergence against the teacher’s temperature-softened soft labels (fixed λ = 0.5, τ = 10).

Load-bearing premise

The same two distillation settings (equal loss weight and temperature 10) are used for every dataset and every architecture, with no per-dataset tuning.

What would settle it

Re-run the same teacher-student pairs on the 112 UCR sets after a modest grid search over λ and τ; if the intermediate-complexity advantage disappears or shifts to other sizes once hyperparameters are tuned, the central pattern does not hold under realistic practice.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies response-based knowledge distillation for univariate time series classification on 112 UCR Archive datasets, using three teacher architectures (FCN, Inception, ConvTran) and families of student models obtained by reducing filters, Inception modules, or attention heads/embedding dimension. Students are trained with a fixed KD loss (λ=0.5, τ=10) against a teacher selected by lowest training loss among five runs, and are compared to matched studentAlone baselines and to the teacher via win/tie/loss counts and accuracy scatter plots. Supporting analyses include filter-space t-SNE for Inception and train/validation loss curves showing reduced generalization gap. The central claim is that KD most benefits intermediate-complexity students: FCN_20F (≈38× fewer parameters), Inception_4M (42% fewer parameters, near-teacher accuracy), and ConvTran_2H (largest student-vs-studentAlone gain).

Significance. If the intermediate-complexity pattern holds under broader hyperparameter and architecture choices, the work provides a useful large-scale empirical baseline for model compression in TSC, where KD has been under-explored relative to vision and NLP. Strengths include evaluation on 112 public datasets, five-run averaging with studentAlone controls, multi-architecture coverage (CNN and transformer), public code, and qualitative analyses (filter transfer, generalization gap) that align with the accuracy results. The contribution is primarily empirical rather than methodological novelty, but the scale and consistency of the findings make it a credible reference for practitioners seeking smaller TSC models.

major comments (2)
  1. Section 4.2 fixes λ=0.5 and τ=10 for every dataset and architecture, citing literature defaults and avoiding test-set tuning. The authors themselves note (discussion of Fig. 5) that fixed hyperparameters may limit KD on the diverse UCR collection. Because the central claim is that intermediate-complexity students benefit most, a minimal sensitivity check (e.g., a small grid of λ,τ on a subset of datasets, or reporting whether the intermediate peak in Figs. 4/6/8 shifts) is needed to show the ranking is not an artifact of this single setting. Without it the claim remains plausible but under-supported as a general pattern.
  2. Results in §§4.3.1–4.3.3 and Figs. 4, 6, 8 rely on raw win counts over 112 datasets with no statistical significance tests (e.g., Wilcoxon signed-rank or critical-difference diagrams standard in TSC). Given five-run averages and many near-ties, it is unclear which win margins are reliable. Adding pairwise significance or CD diagrams would substantially strengthen the intermediate-complexity ranking that carries the paper’s main claim.
minor comments (5)
  1. Table 2 lists the teacher Inception as 422,627 parameters while §3.3.2 states 420,708; reconcile the numbers.
  2. §4.3.2 ends mid-sentence (“However, as shown in Fig. 6b”); complete or remove the fragment.
  3. Fig. 8 axis labels and legend are harder to read than Figs. 4 and 6; improve resolution/labels for consistency.
  4. §4.1.1: briefly justify discarding unequal-length/missing-value datasets and Fungi so the 112-dataset subset is fully reproducible from the text alone.
  5. Related work could more clearly position against Gong et al. (KDCTIME) and the authors’ prior FCN-only study [1] to highlight what is new beyond multi-architecture extension.

Circularity Check

0 steps flagged

No significant circularity: purely empirical win-count and accuracy comparisons of independently trained models on public UCR splits.

full rationale

The paper's central claim (KD most benefits intermediate-complexity students across FCN/Inception/ConvTran) is established solely by experimental measurements: five independent trainings of teacher, student, and studentAlone models per configuration, averaged accuracy, and win/tie/loss counts on the fixed public 112-dataset UCR 2018 splits (Figs. 4–9). No quantity is defined in terms of a fitted parameter later presented as a prediction; λ=0.5 and τ=10 are fixed literature defaults applied uniformly (Sec. 4.2), not optimized on the test outcomes. Filter-space T-SNE (Fig. 10) and loss-curve generalization gaps (Fig. 11) are post-hoc visualizations of already-trained models, not self-referential derivations. Self-citations to the authors' prior FCN-only study [1] merely motivate the extension and are not load-bearing for the new multi-architecture results. The evaluation is therefore self-contained against external benchmarks with matched baselines and public code; no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard KD loss construction, fixed public hyperparameters, public UCR splits, and the modeling choice that students are capacity-reduced copies of the teacher. No new physical entities are introduced; free parameters are the two KD scalars held constant for all experiments.

free parameters (2)
  • distillation weight λ = 0.5
    Set to 0.5 for every dataset and architecture; controls the trade-off between hard-label CE and soft-label KL and is not tuned.
  • softmax temperature τ = 10
    Set to 10 following classic KD literature; softens teacher/student logits and is not tuned per dataset.
axioms (4)
  • domain assumption Response-based KD with KL on temperature-softened logits transfers useful generalization from teacher to student.
    Taken from Hinton et al. and used as the sole distillation mechanism (Eqs. 4–6).
  • domain assumption UCR Archive original train/test splits and z-normalization provide a fair, standard evaluation protocol.
    Stated in Section 4.1.1; 112 of 128 datasets retained after discarding unequal-length, missing-value, and single-example classes.
  • ad hoc to paper Students should be architectural quantizations of the teacher (fewer filters / modules / heads) rather than wholly different designs.
    Explicit design choice in Section 3.4; results are conditional on this family of students.
  • ad hoc to paper Selecting the teacher with lowest training loss among five runs yields a reproducible and representative teacher.
    Section 4.1.2; alternative selection criteria (validation accuracy, ensemble) are not explored.

pith-pipeline@v1.1.0-grok45 · 22344 in / 2640 out tokens · 36337 ms · 2026-07-10T21:09:11.810830+00:00 · methodology

0 comments
read the original abstract

Deep learning has achieved remarkable success in various domains including time series analysis, computer vision and natural language processing. However, high computational and memory demands of state-of-the-art architectures pose challenges for deployment in resource-limited environments. Knowledge Distillation (KD) addresses this by transferring knowledge from a large teacher model to a smaller, more efficient student model while maintaining competitive performance. In this work, we investigate the effectiveness of KD for Time Series Classification (TSC) across three architectures: the classical Fully Convolutional Network (FCN), the convolutional Inception model and the transformer-based ConvTran model. We evaluate our approach on UCR Archive, the largest benchmark repository of time series datasets, by modifying architectural components such as convolutional filters, Inception modules and attention heads across the three architectures. Our results consistently show that KD most effectively benefits student models of intermediate complexity across all three architectures, with the distilled FCN student reducing parameters by a factor of 38, the distilled Inception student achieving nearly the same performance as the teacher with 42% fewer parameters and the distilled ConvTran student with 2 attention heads showing the most significant improvement through distillation. To encourage further research and reproducibility, we provide our implementation at https://github.com/MSD-IRIMAS/KD-4-TSC.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 8 internal anchors

  1. [1]

    International Joint Conference on Neural Networks (IJCNN)

    Ay E, Devanne M, Weber J, Forestier G (2022) A study of knowledge distillation in fully convolutional network for time series classification. International Joint Conference on Neural Networks (IJCNN)

    work page 2022

  2. [2]

    IEEE Trans Knowl Data Eng 270(9):2522–2535

    Bagnall A, Lines J, Hills J, Bostrom A (2015) Time-series classification with cote: the collective of transformation-based ensembles. IEEE Trans Knowl Data Eng 270(9):2522–2535

    work page 2015

  3. [3]

    Data Min Knowl Disc 31:606–660

    Bagnall A, Lines J, Bostrom A, Large J, Keogh E (2017) The great time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Data Min Knowl Disc 31:606–660

    work page 2017

  4. [4]

    Inf Sci 569:596–614

    Baldán FJ, Benítez JM (2021) Multivariate times series classification through an interpretable represen- tation. Inf Sci 569:596–614

    work page 2021

  5. [5]

    IEEE Trans Pattern Anal Mach Intell 350(8):1798–1828

    Bengio Y , Courville A, Vincent P (2013) Representation learning: a review and new perspectives. IEEE Trans Pattern Anal Mach Intell 350(8):1798–1828

    work page 2013

  6. [6]

    Inter- national Conference on Big Data Analytics and Knowledge Discovery

    Bostrom A, Bagnall A (2015) Binary shapelet transform for multiclass time series classification. Inter- national Conference on Big Data Analytics and Knowledge Discovery

    work page 2015

  7. [7]

    In: International conference on Knowledge discovery and data mining

    Buciluˇa C, Caruana R, Niculescu-Mizil A (2006) Model compression. In: International conference on Knowledge discovery and data mining

    work page 2006

  8. [8]

    In: International Conference on Pattern Recognition and Artificial Intelligence (ICPRAI)

    Chelali M, Kurtz C, Puissant A, Vincent N (2020) Spatio-temporal stability analysis in satellite image times series. In: International Conference on Pattern Recognition and Artificial Intelligence (ICPRAI)

    work page 2020

  9. [9]

    IEEE/CAA J Automatica Sinica 60(6):1293–1305

    Dau HA, Bagnall A, Kamgar K, Y eh C-CM, Zhu Y , Gharghabi S, Ratanamahatana CA, Keogh E (2019) The ucr time series archive. IEEE/CAA J Automatica Sinica 60(6):1293–1305

    work page 2019

  10. [10]

    Data Min Knowl Disc 340(5):1454–1495

    Dempster A, Petitjean F, Webb GI (2020) Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Min Knowl Disc 340(5):1454–1495

    work page 2020

  11. [11]

    In: International Conference on Humanoid Robotics (Humanoids)

    Devanne M, Nguyen SM (2017) Multi-level motion analysis for physical exercises assessment in kinaes- thetic rehabilitation. In: International Conference on Humanoid Robotics (Humanoids)

    work page 2017

  12. [12]

    Improving Position Encoding of Transformers for Multivariate Time Series Classification

    Foumani NM, Tan CW, Webb GI, Salehi M (2023) Improving position encoding of transformers for multivariate time series classification. arXiv preprint arXiv:2305.16642

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Inf Sci 613:184–203

    Gong X, Si Y -W, Tian Y , Lin C, Zhang X, Liu X (2022) Kdctime: Knowledge distillation with calibration on inceptiontime for time-series classification. Inf Sci 613:184–203

    work page 2022

  14. [14]

    Int J Comput Vision 129:1789–1819

    Gou J, Baosheng Y u, Maybank SJ, Tao D (2021) Knowledge distillation: a survey. Int J Comput Vision 129:1789–1819

    work page 2021

  15. [15]

    The Twelfth International Conference on Learning Representations

    Gu Y , Dong L, Wei F, Huang M (2024) Minillm: Knowledge distillation of large language models. The Twelfth International Conference on Learning Representations

    work page 2024

  16. [16]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Guo D, Y ang D, Zhang H, Song J, Zhang R, Runxin X, Zhu Q, Ma S, Wang P , Bi X et al (2025) Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Distilling the Knowledge in a Neural Network

    Hinton G, Vinyals O, Dean J (2015) Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  18. [18]

    IEEE J Biomed Health Inform 260(1):423–435

    Ibrahim ZM, Bean D, Searle T, Qian L, Honghan W, Shek A, Kraljevic Z, Galloway J, Norton S, Teo JTH et al (2021) A knowledge distillation ensemble framework for predicting short-and long-term hospi- talization outcomes from electronic health records data. IEEE J Biomed Health Inform 260(1):423–435

    work page 2021

  19. [19]

    In: 2023 IEEE 10th International Conference on Data Science and Advanced Analytics (DSAA), pages 1–10

    Ismail-Fawaz A, Devanne M, Berretti S, Weber J, Forestier G (2023) Lite: Light inception with boosting techniques for time series classification. In: 2023 IEEE 10th International Conference on Data Science and Advanced Analytics (DSAA), pages 1–10. IEEE

    work page 2023

  20. [20]

    Data Min Knowl Disc 330(4):917–963

    Fawaz HI, Forestier G, Weber J, Idoumghar L, Muller P-A (2019) Deep learning for time series classifi- cation: a review. Data Min Knowl Disc 330(4):917–963

    work page 2019

  21. [21]

    Data Min Knowl Disc 340(6):1936–1962

    Fawaz HI, Lucas B, Forestier G, Pelletier C, Schmidt DF, Weber J, Webb GI, Idoumghar L, Muller P-A, Petitjean F, Inceptiontime (2020) Finding alexnet for time series classification. Data Min Knowl Disc 340(6):1936–1962

    work page 2020

  22. [22]

    Data Min Knowl Disc 30:283–312

    Kate RJ (2016) Using dynamic time warping distances as features for improved time series classification. Data Min Knowl Disc 30:283–312

    work page 2016

  23. [23]

    Inf Sci 582:198–214

    Kowsar Y , Moshtaghi M, V elloso E, Bezdek JC, Kulik L, Leckie C (2022) Shape-sphere: A metric space for analysing time series by their shape. Inf Sci 582:198–214

    work page 2022

  24. [24]

    Inf Sci 547:592–608

    Li H (2021) Time works well: Dynamic time warping based on time weighting for time series data mining. Inf Sci 547:592–608

    work page 2021

  25. [25]

    Inf Sci 534:97–116

    Li H, Liu J, Y ang Z, Liu RW, Kefeng W, Wan Y (2020) Adaptively constrained dynamic time warping for time series classification and clustering. Inf Sci 534:97–116

    work page 2020

  26. [26]

    Data Min Knowl Disc 29:565–592

    Lines J, Bagnall A (2015) Time series classification with ensembles of elastic distance measures. Data Min Knowl Disc 29:565–592

    work page 2015

  27. [27]

    International conference on data mining (ICDM) 123 215 Page 22 of 24 J

    Lines J, Taylor S, Bagnall A (2016) Hive-cote: The hierarchical vote collective of transformation-based ensembles for time series classification. International conference on data mining (ICDM) 123 215 Page 22 of 24 J. Abdullayev et al

    work page 2016

  28. [28]

    Pattern Recogn 100:107122

    Mauceri S, Sweeney J, McDermott J (2020) Dissimilarity-based representations for one-class classifica- tion on time series. Pattern Recogn 100:107122

    work page 2020

  29. [29]

    In: International conference on artificial intelligence (AAAI)

    Mirzadeh SI, Farajtabar M, Li A, Levine N, Matsukawa A, Ghasemzadeh H (2020) Improved knowledge distillation via teacher assistant. In: International conference on artificial intelligence (AAAI)

    work page 2020

  30. [30]

    Knowl Inf Syst 410(1):101–125

    Nguyen H-L, Ng W-K, Woon Y -K (2014) Closed motifs for streaming time series classification. Knowl Inf Syst 410(1):101–125

    work page 2014

  31. [31]

    International Joint Conference on Neural Networks (IJCNN)

    Oki H, Abe M, Miyao J, Kurita T (2020) Triplet loss for knowledge distillation. International Joint Conference on Neural Networks (IJCNN)

    work page 2020

  32. [32]

    Knowl Inf Syst 47:1–26

    Petitjean F, Forestier G, Webb GI, Nicholson AE, Chen Y , Keogh E (2016) Faster and more accurate classification of time series by exploiting a novel dynamic time warping averaging algorithm. Knowl Inf Syst 47:1–26

    work page 2016

  33. [33]

    NPJ digital medicine 10(1):18

    Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M et al (2018) Scalable and accurate deep learning with electronic health records. NPJ digital medicine 10(1):18

    work page 2018

  34. [34]

    FitNets: Hints for Thin Deep Nets

    Romero A, Ballas N, Kahou SE, Chassang A, Gatta C, Bengio Y (2014) Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  35. [35]

    Data Min Knowl Disc 29:1505–1530

    Schäfer P (2015) The boss is concerned with time series classification in the presence of noise. Data Min Knowl Disc 29:1505–1530

    work page 2015

  36. [36]

    Shakeri S, Sethy A, Cheng C (2019) Knowledge distillation in document retrieval arXiv preprint arXiv:1911.11065

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  37. [37]

    Patient Knowledge Distillation for BERT Model Compression

    Sun S, Cheng Y u, Gan Z, Liu J (2019) Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  38. [38]

    In: International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Takashima R, Li S, Kawai H (2018) An investigation of a knowledge distillation method for ctc acoustic models. In: International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    work page 2018

  39. [39]

    Knowl Inf Syst 650(11):4967–5004

    Terefe T, Devanne M, Weber J, Hailemariam D, Forestier G (2023) Estimating time series averages from latent space of multi-tasking neural networks. Knowl Inf Syst 650(11):4967–5004. https://doi.org/10. 1007/s10115-023-01981-9

    work page 2023

  40. [40]

    Pattern Recogn 89:55–66

    Wang H, Zhang Q, Jia W, Pan S, Chen Y (2019) Time series feature learning with labeled and unlabeled data. Pattern Recogn 89:55–66

    work page 2019

  41. [41]

    International joint conference on neural networks (IJCNN)

    Wang Z, Y an W, Oates T (2017) Time series classification from scratch with deep neural networks: A strong baseline. International joint conference on neural networks (IJCNN)

    work page 2017

  42. [42]

    Pattern Recogn 111:0–107722

    Wang Z-R, Du J (2021) Joint architecture and knowledge distillation in cnn for chinese text recognition. Pattern Recogn 111:0–107722

    work page 2021

  43. [43]

    Neurocomputing 485:242–251

    Qing X, Chen Z, Ragab M, Wang C, Min W, Li X (2022) Contrastive adversarial knowledge distillation for deep model compression in time-series regression tasks. Neurocomputing 485:242–251

    work page 2022

  44. [44]

    Pattern Recogn 88:272–284

    Ting-Bing X, Y ang P , Zhang X-Y , Liu C-L (2019) Lightweightnet: Toward fast and lightweight convolu- tional neural networks via architecture distillation. Pattern Recogn 88:272–284

    work page 2019

  45. [45]

    A Survey on Knowledge Distillation of Large Language Models

    Xiaohan X, Li M, Tao C, Shen T, Cheng R, Li J, Can X, Tao D, Zhou T (2024) A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [46]

    FITS: Modeling Time Series with $10k$ Parameters

    Zhijian X, Zeng A, Qiang X (2023) Fits: Modeling time series with 10 k parameters. arXiv preprint arXiv:2307.03756

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [47]

    Int J Inf Technol Decis Making 50(04):597–604

    Y ang Q, Xindong W (2006) 10 challenging problems in data mining research. Int J Inf Technol Decis Making 50(04):597–604

    work page 2006

  48. [48]

    In International conference on computer vision and pattern recognition (CVPR)

    Yim J, Joo D, Bae J, Kim J (2017) A gift from knowledge distillation: Fast optimization, network min- imization and transfer learning. In International conference on computer vision and pattern recognition (CVPR)

    work page 2017

  49. [49]

    Zeng A, Chen M, Zhang L, Qiang X (2023) Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence 37:11121–11128

    work page 2023

  50. [50]

    Pattern Recogn 123:108385 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations

    Zhang Y , Hou Y , OuY ang K, Zhou S (2022) Multi-scale signed recurrence plot based time series classi- fication using inception architectural networks. Pattern Recogn 123:108385 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 123 Enhancing deep learning models for time ...

    work page 2022