Pith. sign in

REVIEW 4 major objections 6 minor 55 references

A single causal hidden state can classify a sketch from any unfinished prefix and also generate the next strokes.

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

2026-07-30 18:18 UTC pith:PKPNRRCO

load-bearing objection Clean multi-task recipe for progressive sketch recognition + generation; the dense-loss ablation is real, the Mamba-vs-Transformer parity claim is oversold. the 4 major comments →

arxiv 2607.23580 v1 pith:PKPNRRCO submitted 2026-07-26 cs.CV

SketchMamba: A Lightweight State-Space Model for Joint Progressive Sketch Classification and Stroke Auto-Completion

Kavish Jhaveri , Arya Shah This is my paper
classification cs.CV
keywords vector sketchstate space modelsprogressive recognitionsequence modellingstroke auto-completionQuick Drawcausal Mamba
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.

Streaming sketch apps need to know what you are drawing while you are still drawing it, and to offer stroke completions from the same live state. Prior vector-sketch systems split those jobs: recognisers wait for a finished drawing, generators take class as an external condition. SketchMamba shows that one selective state-space backbone, trained with classification loss at every time step plus ordinary next-stroke losses, can do both from the same causal hidden state. On a 58-class Quick, Draw! subset it reaches 94.93% final accuracy and a progressive-accuracy AUC of 0.706, crossing 90% of final accuracy once about 70% of the strokes are present. A matched-budget comparison ties a causal Transformer and beats recurrent and convolutional baselines; an ablation shows the dense per-step loss, not the backbone alone, is what creates early prediction.

Core claim

Dense per-step classification supervision on a causal selective state-space model forces one hidden state to serve simultaneously as an early-prediction recogniser and as the conditioner for autoregressive stroke generation, without auxiliary encoders or task-specific branching. Final-step accuracy stays high while progressive accuracy rises far earlier than under final-step-only loss.

What carries the argument

Dense per-step classification loss: cross-entropy on the sketch class is averaged over every prefix of the sequence, not only the final token, so the shared causal state must already be class-discriminative before the drawing is complete.

Load-bearing premise

The claim that dense supervision, rather than architecture, drives early prediction rests on one dense-versus-final ablation on the Mamba backbone alone, without progressive-accuracy numbers for the other backbones under the same dense loss.

What would settle it

Train the matched Transformer, LSTM, GRU, and CNN baselines with the identical dense per-step classification loss and report their progressive-accuracy AUC; if they match or exceed 0.706, the supervision-regime claim is supported and architecture is secondary; if they collapse, architecture matters more than the paper concludes.

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

If this is right

  • Streaming ink interfaces can show a live class label and offer stroke auto-completion from one forward pass without recomputing the prefix.
  • Constant-time state updates keep recognition and generation latency flat as the drawing grows, unlike quadratic attention.
  • The same recipe can be tried on larger sketch vocabularies once pen-state imbalance is handled with transition-weighted losses.
  • Progressive-accuracy AUC becomes a practical metric for any causal sketch or handwriting model that claims early understanding.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If dense per-step class loss is the real driver, the same supervision trick should transfer to other linear-time causal backbones (not only selective SSMs) for any prefix-label sequence task.
  • Early geometric ambiguity (circle vs clock, line vs sword) suggests progressive recognition may need explicit multi-hypothesis or hierarchical class heads rather than a single 58-way softmax.
  • Teacher-forced generation metrics leave open whether free-running completions stay class-consistent; a closed-loop sampling study would test true joint competence.

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

4 major / 6 minor

Summary. The manuscript presents SketchMamba, a 1.55M-parameter causal selective state-space (Mamba) model trained on the 58-class Quick, Draw! subset with three per-step heads (58-class classifier, 20-component bivariate GMM stroke head, pen-state head) and a dense per-step classification loss. The paper claims (i) 94.93% final Top-1 accuracy and a progressive-accuracy AUC of 0.706 on the 145k test split, (ii) parity with a causal Transformer at a "matched" parameter budget while beating LSTM/GRU/CNN baselines, and (iii) via ablation, that the dense supervision regime — not the architecture — is the active ingredient for early prediction. The unifying claim is that one causal hidden state supports both streaming recognition and autoregressive stroke generation without auxiliary encoders or task-specific branches.

Significance. If the results hold, the paper makes a useful, if incremental, point for streaming digital-ink interfaces: a single small causal model can deliver per-prefix classification and stroke continuation with O(1) per-step update cost. To its credit, the work is methodologically disciplined in places that matter: all backbones are trained under one fixed pipeline (Table 2), the dense-vs-final-step ablation (Table 4) cleanly separates the supervision effect at matched final accuracy (94.89% vs 94.93%), and the progressive-AUC metric is a concrete, falsifiable measurement on a large held-out test set. The explicit acknowledgment that the supervision regime, rather than the architecture, drives early prediction is refreshingly honest. However, the significance is bounded by the evaluation being confined to a 58-class RDP-simplified Quick, Draw! setting, by the absence of any variance estimates, and by the fact that the central "unification" is demonstrated only for the Mamba backbone, leaving open whether the contribution is the architecture or simply the (known) effectiveness of dense supervision.

major comments (4)
  1. [Table 3 / §5.1] The matched-budget comparison is not actually matched, and is reported without variance. The Transformer baseline carries 1.87M parameters versus 1.55M for SketchMamba (a 20% discrepancy), and the LSTM/GRU/CNN baselines are collapsed into a single row spanning 1.5–1.9M parameters and 88.33–93.43% accuracy, so no per-model comparison is possible. More importantly, the headline parity claim rests on a 0.03-point gap (94.93% vs 94.90%) from what appears to be a single training run per model. Given that the paper itself cites Bouthillier et al. [8] on benchmark variance, single-run numbers at this margin cannot support even the modest claim of a 'tie.' At minimum, report per-model parameter counts and accuracies for each baseline separately, and add multi-seed runs (3 seeds would suffice) with standard deviations for SketchMamba and the Transformer.
  2. [Table 3, Prog. AUC column] Progressive AUC is reported as '–' for every baseline. This omission is load-bearing: §4.2 states that all baselines share 'the exact same embedding layer, task heads, and training hardware' under the fixed pipeline, which means the baselines were trained with the same dense per-step classification loss and therefore have well-defined progressive curves that are simply not reported. Since the paper's own Table 4 ablation attributes progressive recognition to the loss rather than the architecture, the natural reading is that a Transformer or LSTM trained with the dense loss would achieve comparable AUC — in which case the 'unification' capability is not specific to the state-space backbone and the framing of the paper (and its title) overstates what is demonstrated. Reporting the baselines' progressive AUCs, which requires no retraining, would either substantiate or appropriately scope th
  3. [§5.3 vs §4.3] Generation quality is asserted but never quantified. §4.3 promises that 'generation fidelity is measured via the negative log-likelihood (NLL) of the stroke spatial coordinates and pen state predictions,' yet §5.3 reports no NLL numbers whatsoever — only the qualitative statement that the model generates 'semantically appropriate strokes' plus latency measurements. For a paper whose title and abstract claim joint classification and stroke auto-completion as co-equal contributions, the generation half of the claim is currently unsupported by any reported metric, sample figure, or comparison (e.g., against a Sketch-RNN-style decoder at matched budget). Please report the promised NLL values for SketchMamba and the baselines, and ideally qualitative continuation samples from partial prefixes.
  4. [§5.2 / Table 4] The dense-vs-final-step ablation is cleanly executed but its interpretation should be tempered. A classification head that receives no gradient at early steps trivially cannot classify at early steps, so the AUC collapse (0.706 → 0.074) confirms necessity of the loss almost by construction; it does not, as §5.2 currently implies, reveal something non-obvious about how dense supervision 'structures the intermediate hidden states.' The result is worth keeping, but the discussion should acknowledge the near-tautological nature of the finding and, ideally, be complemented by a more informative variant — e.g., partial-density supervision (classify only after fraction r of the sequence) — to show where early predictability actually emerges.
minor comments (6)
  1. [§5.4 / Fig. 4] The confusion-matrix discussion cites 'lines versus swords' as a confusable pair, but neither 'line' nor 'sword' appears in the 58-class list shown in Fig. 4's axis labels. Please correct the examples to classes actually in the subset.
  2. [Fig. 2] The caption states the figure compares 'dense vs. final-step supervision,' but the plotted legend shows only train/val curves. Either plot both supervision regimes or fix the caption.
  3. [Fig. 3 / §5.3] The latency curve rises visibly from ~1.5 ms to ~2.0 ms across sequence lengths 50–200, which is in tension with the strict 'O(1)' and 'constant irrespective of prefix length' wording; the measured growth is small and practically irrelevant, but the text should describe the figure accurately. Relatedly, the claim of 'bypassing the O(L) scaling overhead of the Transformer' is made without any Transformer latency measurement — either measure it or soften the claim.
  4. [§2.4 / §4.3] Macro-F1 is motivated in §2.4 for class imbalance but never reported; either report it or drop the discussion.
  5. [General] No code or checkpoint release is mentioned. Given that reproducibility (fixed pipeline, Table 2 hyperparameters) is a stated methodological value of the paper, releasing training/evaluation code would substantially strengthen the submission.
  6. [§5.4] The pen-state imbalance discussion is welcome, but a quantitative number (e.g., rate of non-terminating samples) would make the limitation concrete rather than anecdotal.

Circularity Check

0 steps flagged

No circularity: standard supervised training and held-out evaluation; progressive AUC and dense-vs-final ablation are measured outputs, not forced by construction.

full rationale

SketchMamba’s load-bearing claims are empirical ML results: a causal Mamba backbone trained with dense per-step classification plus next-stroke GMM/pen losses achieves 94.93% final Top-1 and progressive-accuracy AUC 0.706 on a held-out 58-class Quick, Draw! split, with an ablation showing final-step-only loss collapses progressive AUC to 0.074, and a matched-budget comparison tying a causal Transformer. None of these quantities is defined in terms of the claimed result, fitted then re-reported as a prediction, or justified by a self-citation uniqueness theorem. Labels and next-token targets come from the external Quick, Draw! corpus; progressive accuracy is read from model logits on test prefixes; the ablation compares two training objectives on the same backbone. Citations (Mamba, Sketch-RNN GMM head, Transformer, etc.) are standard external priors, not author-overlapping load-bearing uniqueness results. Methodological weaknesses (missing progressive AUC for baselines, single-run 0.03-point parity, somewhat tautological necessity of early supervision) affect correctness risk, not circularity. The derivation chain is ordinary supervised learning against external benchmarks and is self-contained.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

Empirical deep-learning paper. Load-bearing content is architectural and training choices plus public sketch labels, not new physical entities. Free parameters are the usual optimizer/architecture/loss knobs that define the reported operating point. Axioms are standard sequence-modeling and dataset assumptions. No graviton-like invented mediators.

free parameters (5)
  • loss weights (λ_cls, λ_stroke, λ_pen) = (1.0, 1.0, 0.5)
    Hand-set multi-task weights that balance recognition vs generation; central joint-training claim depends on this operating point.
  • GMM component count K = 20
    Number of bivariate mixture components in the stroke head; controls generative capacity and NLL.
  • backbone width/depth and matched param budget ~1.5–1.9M = Mamba 1.55M; Transformer 1.87M; others 1.5–1.9M
    Architecture sizes chosen to compare Mamba vs Transformer/LSTM/GRU/CNN; absolute accuracy depends on these choices.
  • peak LR, warmup, weight decay, EMA, epochs, batch size = LR 5e-4, 5000 warmup, wd 0.1, EMA 0.9999, 50 epochs, bs 1024
    Training recipe held fixed across backbones; recipe can dominate architecture deltas as the paper itself notes via citations.
  • RDP epsilon and max length T = ε=2.0, T=200
    Preprocessing choices that define the sequence distribution the metrics are computed on.
axioms (5)
  • domain assumption Quick, Draw! crowd labels and the chosen 58-class subset are a valid proxy for streaming sketch recognition difficulty.
    All accuracy and progressive AUC claims are measured only on this labeled subset (§4.1).
  • domain assumption Teacher-forced next-step GMM/pen training plus causal state updates transfers to true user-in-the-loop streaming inference.
    Streaming inference is described in §3.4 but experiments report teacher-forced/offline prefix metrics, not interactive user studies.
  • domain assumption Selective state-space recurrence provides a strictly causal hidden state h_t depending only on x_1..t suitable for joint heads.
    Invoked throughout §3.2–3.3 via Mamba/S4 lineage citations; standard SSM assumption.
  • ad hoc to paper Matched parameter count under one fixed training pipeline is a fair primary comparison across LSTM/GRU/CNN/Transformer/Mamba.
    Stated motivation in §1 and §4.2; fairness still depends on untuned-per-architecture hyperparameters.
  • standard math Standard cross-entropy, AdamW, cosine schedule, RMSNorm/bf16 numerics are appropriate without further justification.
    Training protocol §4.3 / Table 2 follows cited common practice.
invented entities (1)
  • SketchMamba (joint dense-supervised causal Mamba sketch model) no independent evidence
    purpose: Name the proposed end-to-end system combining Mamba backbone, three per-step heads, and dense classification loss.
    System name for a composition of existing components rather than a new physical or mathematical object; independent evidence is the reported empirical metrics only.

pith-pipeline@v1.2.0-grok45-kimik3 · 15378 in / 3775 out tokens · 76227 ms · 2026-07-30T18:18:32.691663+00:00 · methodology

0 comments
read the original abstract

Existing vector-sketch models treat recognition and generation as separate tasks, leaving a gap for streaming interfaces that must understand a drawing as it is being made. We present SketchMamba, a single causal sequence model that continuously classifies a sketch from any partial prefix while simultaneously generating its continuation. We achieve this by applying a dense per-step classification loss to a selective state-space backbone. Evaluated on a 58-class subset of the Quick, Draw! dataset, SketchMamba yields 94.93% final-step accuracy and a progressive-accuracy Area Under the Curve (AUC) of 0.706, crossing 90% of its final accuracy by the time 70% of the strokes are drawn. In a matched-budget comparison, the 1.55 million-parameter backbone ties a causal Transformer while outperforming recurrent and convolutional baselines. Ablations confirm that the dense supervision regime, rather than the architecture alone, drives the early-prediction capability. The results demonstrate that a single causal hidden state can unify progressive recognition and autoregressive generation without auxiliary encoders or task-specific branching.

Figures

Figures reproduced from arXiv: 2607.23580 by Arya Shah, Kavish Jhaveri.

Figure 1
Figure 1. Figure 1: Progressive classification on the 145,000-sketch Quick, Draw! test split (58 classes). SketchMamba reaches 75.8% Top-1 by half a sketch and over 91% Top-5 by 40% of the strokes; the area under this curve, normalised to the unit interval, is 0.706. The curve is read at any prefix length r from the same causal hidden state used to generate stroke continuations. classify finished sketches, leaving per-prefix … view at source ↗
Figure 2
Figure 2. Figure 2: Training loss curves comparing dense vs. final-step supervision. 50 75 100 125 150 175 200 Sequence length 1.5 1.6 1.7 1.8 1.9 2.0 Latency (ms / sequence, GPU) Forward-pass latency [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Confusion matrix detailing early-stroke ambiguities. 5.4 Discussion and Limitations Analysis of the confusion matrix ( [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

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

55 extracted references · 3 canonical work pages

  1. [1]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Aksan, E., Deselaers, T., Tagliasacchi, A., Hilliges, O.: Cose: Compositional stroke embeddings. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 10041–10052. Curran Associates, Inc. (2020)

    2020

  2. [2]

    ArXivabs/1803.01271 (2018)

    Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. ArXivabs/1803.01271 (2018)

    Pith/arXiv arXiv 2018

  3. [3]

    ArXivabs/2103.07579 (2021)

    Bello, I., Fedus, W., Du, X., Cubuk, E.D., Srinivas, A., Lin, T.Y., Shlens, J., Zoph, B.: Revisiting resnets: Improved training and scaling strategies. ArXivabs/2103.07579 (2021)

    Pith/arXiv arXiv 2021

  4. [4]

    ArXiv abs/2004.05150 (2020)

    Beltagy, I., Peters, M.E., Cohan, A.: Longformer: The long-document transformer. ArXiv abs/2004.05150 (2020)

    Pith/arXiv arXiv 2004

  5. [5]

    In: Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVII

    Bhunia, A.K., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S., Laaksonen, J., Felsberg, M.: Doodleformer: Creative sketch drawing with transformers. In: Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVII. p. 338–355. Springer-Verlag, Berlin, Heidelberg (2022). doi:10.1007/978-3-031-19790-1_21

    work page doi:10.1007/978-3-031-19790-1_21 2022

  6. [6]

    so you think you can sketch? ACM Trans

    Bhunia, A.K., Das, A., Muhammad, U.R., Yang, Y., Hospedales, T.M., Xiang, T., Gryaditskaya, Y., Song, Y.Z.: Pixelor: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph. 39(6) (Nov 2020). doi:10.1145/3414685.3417840

    arXiv 2020

  7. [7]

    WorkingPaper 4288, Aston University (1994), copyright © 1994, Christopher M

    Bishop, C.: Mixture density networks. WorkingPaper 4288, Aston University (1994), copyright © 1994, Christopher M. Bishop. This work is licensed under a Cre- ative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (https://creativecommons.org/licenses/by-nc-nd/4.0/)

    1994

  8. [8]

    ArXiv abs/2103.03098 (2021)

    Bouthillier, X., Delaunay, P., Bronzi, M., Trofimov, A., Nichyporuk, B., Szeto, J., Sepah, N., Raff, E., Madan, K., Voleti, V.S., Kahou, S.E., Michalski, V., Serdyuk, D., Arbel, T., Pal, C.J., Varoquaux, G., Vincent, P.: Accounting for variance in machine learning benchmarks. ArXiv abs/2103.03098 (2021)

    Pith/arXiv arXiv 2021

  9. [9]

    In: Proceedings of the 34th International Conference on Neural Information Processing Systems

    Carlier, A., Danelljan, M., Alahi, A., Timofte, R.: Deepsvg: a hierarchical generative network for vector graphics animation. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20, Curran Associates Inc., Red Hook, NY, USA (2020)

    2020

  10. [10]

    Machine Learning28, 41–75 (1997)

    Caruana, R.: Multitask learning. Machine Learning28, 41–75 (1997)

    1997

  11. [11]

    ArXiv abs/1709.04121 (2017)

    Chen, Y., Tu, S., Yi, Y., Xu, L.: Sketch-pix2seq: a model to generate sketches of multiple categories. ArXiv abs/1709.04121 (2017)

    Pith/arXiv arXiv 2017

  12. [12]

    In: Conference on Empirical Methods in Natural Language Processing (2014) SketchMamba 11

    Cho, K., van Merrienboer, B., Çaglar Gülçehre, Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder–decoder for statistical machine translation. In: Conference on Empirical Methods in Natural Language Processing (2014) SketchMamba 11

    2014

  13. [13]

    In: Korhonen, A., Traum, D., Màrquez, L

    Dai, Z., Yang, Z., Yang, Y., Carbonell, J., Le, Q., Salakhutdinov, R.: Transformer-XL: Atten- tive language models beyond a fixed-length context. In: Korhonen, A., Traum, D., Màrquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Lin- guistics. pp. 2978–2988. Association for Computational Linguistics, Florence, Ital...

    doi:10.18653/v1/p19-1285 2019

  14. [14]

    ArXivabs/2212.14052 (2022)

    Dao, T., Fu, D.Y., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. ArXivabs/2212.14052 (2022)

    Pith/arXiv arXiv 2022

  15. [15]

    ArXivabs/2405.21060 (2024)

    Dao, T., Gu, A.: Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. ArXivabs/2405.21060 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    In: International Conference on Learning Representations (2022)

    Das, A., Yang, Y., Hospedales, T., Xiang, T., Song, Y.Z.: SketchODE: Learning neural sketch representation in continuous time. In: International Conference on Learning Representations (2022)

    2022

  17. [17]

    In: European Conference on Computer Vision (2020)

    Das, A., Yang, Y., Hospedales, T.M., Xiang, T., Song, Y.Z.: Béziersketch: A generative model for scalable vector sketches. In: European Conference on Computer Vision (2020)

    2020

  18. [18]

    In: International Con- ference on Learning Representations (2021)

    Ge, S., Goswami, V., Zitnick, L., Parikh, D.: Creative sketch generation. In: International Con- ference on Learning Representations (2021)

    2021

  19. [19]

    ArXivabs/1308.0850 (2013)

    Graves, A.: Generating sequences with recurrent neural networks. ArXivabs/1308.0850 (2013)

    Pith/arXiv arXiv 2013

  20. [20]

    In: First Conference on Language Modeling (2024)

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. In: First Conference on Language Modeling (2024)

    2024

  21. [21]

    ArXiv abs/2008.07669 (2020)

    Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimal poly- nomial projections. ArXiv abs/2008.07669 (2020)

    Pith/arXiv arXiv 2008

  22. [22]

    ArXiv abs/2111.00396 (2021)

    Gu, A., Goel, K., R’e, C.: Efficiently modeling long sequences with structured state spaces. ArXiv abs/2111.00396 (2021)

    Pith/arXiv arXiv 2021

  23. [23]

    arXiv preprint arXiv:1704.03477 (2017)

    Ha, D., Eck, D.: A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477 (2017)

    Pith/arXiv arXiv 2017

  24. [24]

    Neural Comput

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780 (Nov 1997). doi:10.1162/neco.1997.9.8.1735

    doi:10.1162/neco.1997.9.8.1735 1997

  25. [25]

    In: Conference on Uncertainty in Artificial Intelli- gence (2018)

    Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D.P., Wilson, A.G.: Averaging weights leads to wider optima and better generalization. In: Conference on Uncertainty in Artificial Intelli- gence (2018)

    2018

  26. [26]

    Jongejan, J., Rowley, H., Kawashima, T., Kim, J., Fox-Gieg, N.: The quick, draw! dataset (2017)

    2017

  27. [27]

    ArXivabs/1905.12322 (2019)

    Kalamkar, D.D., Mudigere, D., Mellempudi, N., Das, D., Banerjee, K., Avancha, S., Vooturi, D.T., Jammalamadaka, N., Huang, J., Yuen, H., Yang, J., Park, J., Heinecke, A., Georganas, E., Srinivasan, S.M., Kundu, A., Smelyanskiy, M., Kaul, B., Dubey, P.K.: A study of bfloat16 for deep learning training. ArXivabs/1905.12322 (2019)

    Pith/arXiv arXiv 1905

  28. [28]

    In: International Conference on Machine Learning (2020)

    Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: International Conference on Machine Learning (2020)

    2020

  29. [29]

    IEEE Transactions on Visualization and Computer Graphics pp

    Li, L., Zou, C., Zheng, Y., Su, Q., Fu, H., Tai, C.L.: Sketch-R2CNN: An rnn-rasterization-cnn architecture for vector sketch recognition. IEEE Transactions on Visualization and Computer Graphics pp. 1–1 (2020). doi:10.1109/TVCG.2020.2987626

    arXiv 2020

  30. [30]

    2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Lin, H., Fu, Y., Jiang, Y.G., Xue, X.: Sketch-bert: Learning sketch bidirectional encoder rep- resentation from transformers by self-supervised learning of sketch gestalt. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6757–6766 (2020)

    2020

  31. [31]

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11966–11976 (2022)

    2022

  32. [32]

    arXiv: Learning (2016)

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning (2016)

    2016

  33. [33]

    In: International Conference on Learning Representations (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2017)

    2017

  34. [34]

    In: Speech Synthesis Workshop (2016)

    van den Oord, A., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A.W., Kavukcuoglu, K.: Wavenet: A generative model for raw audio. In: Speech Synthesis Workshop (2016)

    2016

  35. [35]

    In: Conference on Empirical Methods in Natural Language Processing (2023)

    Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., Kranthikiran, G., Du, X., He, X., Hou, H., Kazienko, P., Kocoń, J., Kong, J., Koptyra, B., Lau, H., Mantri, K.S.I., Mom, F., Saito, A., Tang, X., Wang, B., Wind, J.S., Wozniak, S., Zhang, R., Zhang, Z., Zhao, Q., Zhou, P., Zhu, J., Zhu,...

    2023

  36. [36]

    In: International Confer- ence on Machine Learning (2023)

    Poli, M., Massaroli, S., Nguyen, E., Fu, D.Y., Dao, T., Baccus, S.A., Bengio, Y., Ermon, S., Ré, C.: Hyena hierarchy: Towards larger convolutional language models. In: International Confer- ence on Machine Learning (2023)

    2023

  37. [37]

    2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp

    Qi, Y., Su, G., Chowdhury, P.N., Li, M., Song, Y.Z.: Sketchlattice: Latticed representation for sketch manipulation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 933–941 (2021)

    2021

  38. [38]

    Qi, Y., Su, G., Wang, Q., Yang, J., Pang, K., Song, Y.Z.: Generative sketch healing. Int. J. Comput. Vision 130(8), 2006–2021 (Aug 2022). doi:10.1007/s11263-022-01623-7 12 K. Jhaveri et al

    work page doi:10.1007/s11263-022-01623-7 2006

  39. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ribeiro, L.S.F., Bui, T., Collomosse, J., Ponti, M.: Sketchformer: Transformer-based represen- tation for sketched structure. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14153–14162 (2020)

    2020

  40. [40]

    ArXivabs/1706.05098 (2017)

    Ruder, S.: An overview of multi-task learning in deep neural networks. ArXivabs/1706.05098 (2017)

    Pith/arXiv arXiv 2017

  41. [41]

    ACM Transactions on Graphics (TOG) 35, 1 – 12 (2016)

    Sangkloy, P., Burnell, N., Ham, C., Hays, J.: The sketchy database. ACM Transactions on Graphics (TOG) 35, 1 – 12 (2016)

    2016

  42. [42]

    In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

    Sivaprasad, P.T., Mai, F., Vogels, T., Jaggi, M., Fleuret, F.: Optimizer benchmarking needs to account for hyperparameter tuning. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 9036–9045. PMLR (2020)

    2020

  43. [43]

    ArXiv abs/2208.04933 (2022)

    Smith, J., Warrington, A., Linderman, S.W.: Simplified state space layers for sequence modeling. ArXiv abs/2208.04933 (2022)

    Pith/arXiv arXiv 2022

  44. [44]

    Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for classification tasks. Inf. Process. Manag.45, 427–437 (2009)

    2009

  45. [45]

    Su, G., Qi, Y., Pang, K., Yang, J., Song, Y.Z.: Sketchhealer a graph-to-sequence network for recreating partial human sketches (20200907 - 20200910)

  46. [46]

    ArXivabs/2307.08621 (2023)

    Sun, Y., Dong, L., Huang, S., Ma, S., Xia, Y., Xue, J., Wang, J., Wei, F.: Retentive network: A successor to transformer for large language models. ArXivabs/2307.08621 (2023)

    Pith/arXiv arXiv 2023

  47. [47]

    In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polo- sukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)

    2017

  48. [48]

    ACM Transactions on Graphics (TOG)41, 1 – 11 (2022)

    Vinker, Y., Pajouheshgar, E., Bo, J.Y., Bachmann, R., Bermano, A.H., Cohen-Or, D., Zamir, A., Shamir, A.: Clipasso. ACM Transactions on Graphics (TOG)41, 1 – 11 (2022)

    2022

  49. [49]

    In: The Eleventh International Conference on Learning Representations (2023)

    Wang, Q., Deng, H., Qi, Y., Li, D., Song, Y.Z.: Sketchknitter: Vectorized sketch generation with diffusion models. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  50. [50]

    Xing, Z., Pei, J., Yu, P.S.: Early classification on time series. Knowl. Inf. Syst.31(1), 105–127 (Apr 2012)

    2012

  51. [51]

    2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp

    Xu, P., Huang, Y., Yuan, T., Pang, K., Song, Y.Z., Xiang, T., Hospedales, T.M., Ma, Z., Guo, J.: Sketchmate: Deep hashing for million-scale human sketch retrieval. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 8090–8098 (2018)

    2018

  52. [52]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Yang, L., Pang, K., Zhang, H., Song, Y.Z.: Sketchaa: Abstract representation for abstract sketches. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10097–10106 (October 2021)

    2021

  53. [53]

    ACM Trans

    Yang, L., Zhuang, J., Fu, H., Wei, X., Zhou, K., Zheng, Y.: Sketchgnn: Semantic sketch segmen- tation with graph neural networks. ACM Trans. Graph.40(3) (Aug 2021). doi:10.1145/3450284

    work page doi:10.1145/3450284 2021

  54. [54]

    In: British Machine Vision Conference (2015)

    Yu, Q., Yang, Y., Song, Y.Z., Xiang, T., Hospedales, T.M.: Sketch-a-net that beats humans. In: British Machine Vision Conference (2015)

    2015

  55. [55]

    ArXivabs/1910.07467 (2019)

    Zhang, B., Sennrich, R.: Root mean square layer normalization. ArXivabs/1910.07467 (2019)

    Pith/arXiv arXiv 1910