Pith. sign in

REVIEW 2 major objections 5 minor 60 references

A spiking neural transformer recognizes handwritten text in just two timesteps, reaching test character error rates of 5.4% on IAM, 2.5% on LAM, and 3.9% on READ2016, without pretraining, language models, or lexicons.

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 · deepseek-v4-flash

2026-08-04 23:36 UTC pith:7SK2Z5GX

load-bearing objection Solid, well-ablated empirical paper; accuracy is credible, but the spiking-specific efficiency benefit is unproven and the real load-bearing gain is the ANN-compatible length reducer. the 2 major comments →

arxiv 2608.01646 v1 pith:7SK2Z5GX submitted 2026-08-03 cs.NE

Spike-HTR: Spiking Neural Transformer for Handwritten Text Recognition

Xiubo Liang , Jinxing Han , Yuke Li , Haoqi Zhu , Yu Zhao , Hongzhi Wang This is my paper
classification cs.NE
keywords spiking neural networkshandwritten text recognitionspiking transformerCTC length reductioncoarse-to-fine temporal codingcharacter error rateneuromorphic efficiencyblank-posterior pruning
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.

This paper tries to establish that a spiking neural network can recognize offline handwriting at near-conventional accuracy with only two simulation timesteps, if computation is budgeted in two places: time and sequence length. The temporal budget is spent as coarse-to-fine refinement — InkCoder converts the static line image into deterministic gates so early timesteps admit broad stroke support and later timesteps sharpen edge detail — rather than replaying the same frame. The sequence budget is spent selectively: a CTC-guided reducer keeps likely character and uncertain positions and compresses long blank-dominated stretches before the deep mixer, shortening token mixing by about 32% with negligible character error change. Trained from scratch on each target dataset and decoded with greedy CTC and no language model, Spike-HTR reaches test character error rates of 5.4% on IAM, 2.5% on LAM, and 3.9% on READ2016. A careful reader should care because static handwriting is a poor native fit for time-unrolled spiking computation, and the paper shows the mismatch can be engineered away with two explicit budgets rather than longer simulation.

Core claim

Spike-HTR treats offline handwriting recognition as a two-budget allocation problem. The temporal budget is spent as coarse-to-fine refinement: InkCoder converts the static line image into deterministic gates so early spiking steps admit broad stroke support and later steps sharpen edge detail. The sequence budget is spent selectively: a CTC-guided reducer keeps likely character and uncertain positions and compresses long blank runs before deep mixing, shortening the token sequence by about 32% with negligible CER change. With T=2, trained from scratch and decoded greedily without language models, Spike-HTR reaches test CERs of 5.4 on IAM, 2.5 on LAM, and 3.9 on READ2016, and its validation

What carries the argument

The central mechanisms are two budget controllers. InkCoder turns one static line image into T deterministic spatial gates: it fuses an intensity proxy that captures stable stroke mass with a coherence-filtered edge proxy that preserves thin strokes and junctions, under a rising threshold schedule so early timesteps admit broad stroke support and later timesteps sharpen edge detail; the gates modulate a shared non-spiking stem feature fed into the spiking 2D encoder. The CTC-guided length reducer runs a stop-gradient blank-posterior preview through the shared classification head, keeps positions with low blank posterior or high uncertainty, collapses long blank-like runs into bounded spans,

Load-bearing premise

The efficiency payoff assumes that sparse spike activity turns into proportional energy and memory savings on real hardware; the paper's support for that is an idealized arithmetic estimate that ignores data movement, and on today's dense accelerators the certain gain is only the one-third sequence shortening, not the spiking pathway itself.

What would settle it

Run the T=2 model on an event-driven neuromorphic or structured-sparsity backend and measure per-image energy against a dense-arithmetic baseline at matched accuracy: the paper's proxy predicts energy that scales with the 32% sequence reduction and the measured firing rates (median 2.18%), so if observed savings track only the sequence shortening — or none at all — the spiking-specific efficiency claim collapses. A second check: run the reducer on a densely spaced, blank-free text layout; if the 32% saving does not shrink or CER degrades, the reduction is exploiting blank redundancy rather tha

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A spiking recognizer with two timesteps can match conventional recognizers on line-level handwriting, so static-image spiking inference does not need long temporal simulation when the input is structured as a coarse-to-fine stream.
  • Because the reducer needs only a CTC head, the keep-and-merge length control transfers to ordinary ANN-CTC recognizers, offering about a 32% shorter token-mixing sequence with negligible character error change.
  • Timesteps beyond T=2 buy word-level corrections rather than character accuracy: test WER drops from 19.4 to 18.8 on IAM, 9.7 to 8.5 on LAM, and 20.2 to 19.7 on READ2016 at T=4, so users can trade temporal compute for WER.
  • Spike-volume profiling shows early 2D encoder layers emit 96.65% of all spikes despite lower firing rates, so further efficiency work should target the early visual pathway and structured-sparsity execution rather than the already-compact mixer.

Where Pith is reading between the lines

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

  • Inference beyond the paper: the same two-budget recipe should transfer to other CTC-trained sequence tasks with blank-heavy alignments, such as speech-to-text frames or audio event detection, where a stop-gradient blank-posterior preview could compress stretches of silence or background before deep mixing.
  • Inference beyond the paper: the energy story is only as strong as the hardware assumption, so the decisive experiment is measuring the T=2 model on a real event-driven backend; the paper's arithmetic proxy assumes fixed per-operation costs and ignores memory traffic, so a measured energy-vs-dense baseline would separate the spiking-specific gain from the sequence-shortening gain.
  • Inference beyond the paper: InkCoder's gates are built from generic intensity/edge/coherence priors, so they should port to other static-image spiking tasks such as document binarization or signature verification; a testable extension is to learn the gate schedule per domain rather than fixing it and check whether the T=2 saturation still holds.

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 proposes Spike-HTR, a hybrid spiking-transformer recognizer for offline handwritten text recognition. A static line image is converted into a short coarse-to-fine temporal stream by InkCoder, and a CTC-guided length reducer prunes/merges blank-dominated width-axis positions before a deep 1D mixer. With T=2, no external pretraining, no language model, and greedy CTC decoding, the method reports test CERs of 5.4, 2.5, and 3.9 on IAM, LAM, and READ2016. Extensive ablations isolate the contributions of InkCoder, the reduction rule, the spiking 2D encoder, and the 1D mixer; a matched ANN-isomorphic control is reported on IAM; and spike-activity profiling plus a compute-only arithmetic proxy are used to discuss efficiency.

Significance. If the efficiency claims held, this would be a useful contribution: it shows that a spiking transformer can reach competitive HTR accuracy with very few timesteps and that CTC-driven sequence reduction yields a 32% token-mixing saving. The experimental breadth is a real strength: multiple datasets, reduction ablations, InkCoder ablations, encoder/mixer ablations, a T-sweep, and a matched ANN control on IAM. The paper is also honest about many limitations and releases code, which supports reproducibility. The accuracy story is credible. The efficiency story, however, is not established for spiking-specific execution; the only directly measured saving is the ANN-compatible length-reduction effect.

major comments (2)
  1. [§4.3 / D.3 / Table 18 / Section E] The spiking-specific efficiency contribution is not established. The compute-only proxy (Eqs. 33–36) ignores memory traffic, data movement, and real accelerator behavior; even under this favorable proxy, the default configuration in Table 18 gives E_hyb/E_dense = 0.642 with only 35.9% of arithmetic event-driven-eligible. The pure-spike variant (0.525) conflicts with the accuracy ablation in Table 14, since removing the analog bypass raises CER on all datasets, and the ideal 0.00495 is explicitly not an implementation result. Section E concedes that on current dense accelerators the immediate algorithmic benefit is the width reduction, which is ANN-compatible (Table 5, r≈0.68). Thus "cutting token-mixing computation" is supported as a length-reduction effect, but the spiking-specific energy/sparsity advantage remains a hypothesis. Please provide memory-aware or hardware measurements, or e
  2. [§4.1, Table 3 vs Table 1] The claim of near-ANN accuracy over all three datasets rests on a matched ANN control that is only reported for IAM validation (Table 3), and the differences there are small (3.489 vs 3.558 vs 3.646) with no variance or multiple-seed statistics. For LAM and READ2016, Table 1 compares against prior ANN systems whose training data, pretraining, and decoding protocols differ, as the caption concedes. To support near-ANN accuracy as a general claim, please add matched ANN baselines for LAM and READ2016, or restrict the claim to IAM and report seed statistics.
minor comments (5)
  1. [§3.2 / Eq. (32)] The text says InkCoder learns only "two global gate-sharpness scalars," but Eq. (32) lists several schedule parameters (a0, η_a, s_α, b_α, s_w, b_w). Clarify which parameters are learned, which are fixed, and the actual total parameter count.
  2. [Table 1] The table formatting for several rows appears shifted, with extra or missing numeric columns. Please reformat so that IAM, LAM, and READ2016 columns align correctly.
  3. [Table 5 vs Table 14] The full-model validation CER is 3.49 in Table 5 but 3.4 in Table 14 under what appears to be the same base T=2 configuration. Reconcile the numbers or clarify the difference.
  4. [Appendix D.4 / Table 18] Define "eligible share" explicitly; from the numbers it appears to be (N_dense − N_hyb_MAC)/N_dense, but the current text leaves this to inference.
  5. [Eq. (19)] The KL denominator divides by the sum of valid-position masks. If a degenerate padded batch had all masks zero this would be undefined; add a small epsilon or note that valid lengths are always positive.

Circularity Check

0 steps flagged

No significant circularity: the accuracy results are external benchmark comparisons and the efficiency numbers are measured diagnostics, not fitted predictions.

full rationale

Spike-HTR is an empirical system paper. The central accuracy claim (3.5/5.4, 2.3/2.5, 4.2/3.9 CER) is obtained by training from scratch on each target dataset and evaluating greedy CTC on held-out test splits; no target quantity is built into the training objective or defined in terms of the reported result. The reduction ratio r=0.68 is a measured statistic from the trained checkpoint, not a parameter fitted to the CER change; the ablation 'No reduction' provides an independent control. The efficiency analysis (Appendix D.3) is explicitly a compute-only arithmetic proxy with externally stated per-operation costs (45 nm pJ values), and the paper reports measured firing rates rather than assuming them; even the 'ideal' bound is labeled 'not an implementation result.' The paper itself concedes in Section E that on dense accelerators the immediate benefit is the width-axis length reduction and that spiking-specific gains require event-driven hardware/sparse kernels. That is a limitation of evidence, not a circular reduction. The only candidate self-citations (Liang et al. 2022 on token pruning; Li et al. 2025 HTR-VT used as a baseline in Table 1) are related-work context and are not load-bearing: the paper's conclusions do not rest on those papers. No uniqueness theorem, ansatz-by-citation, or self-definitional equation was found. Therefore no circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central accuracy results rest on standard CTC training and surrogate-gradient SNN training (axioms 1-2), plus domain assumptions that the handwriting priors encoded in InkCoder and the blank-posterior preview are informative for the three Latin-script benchmarks (axioms 3-4). The efficiency claim additionally rests on a compute proxy that the paper itself labels idealized (axiom 5). Free parameters are limited to fixed thresholds and schedule scalars; two gate-sharpness scalars are learned. No new physical or architectural entities are introduced beyond the InkCoder gates.

free parameters (6)
  • blank keep threshold tau = 0.88
    Hand-set threshold for blank posterior below which a position is kept in the CTC-guided reduction (Table 8); affects reduction ratio r and CER.
  • minimum keep ratio gamma = 0.70
    Hand-set minimum fraction of positions that must be retained per sample (Eq. 16 and Table 8); controls the feasible reduction budget.
  • merge span cap k = 3
    Maximum size of merged blank-like spans (Section 3.4, Table 8); affects how aggressively long blanks are compressed.
  • temporal injection baseline beta = 0.35
    Non-zero baseline in Eq. 7 to prevent a fully silent drive under aggressive gating; hand-set in base config.
  • InkCoder gate schedule parameters = theta_min, theta_max, gamma_theta, a_0, b_a, s_w, b_w
    Hand-crafted monotone threshold and sharpness schedules (Eq. 32) that define the coarse-to-fine gate; only two gate-sharpness scalars are learned.
  • auxiliary loss weights lambda_aux, lambda_kl = 0.2, 0.05
    Weights of auxiliary CTC and cross-depth KL losses (Eq. 19) chosen in the base training recipe (Table 10).
axioms (5)
  • standard math Multi-step LIF with surrogate gradients is a valid differentiable approximation for training spiking networks.
    Used throughout Section 3.1 and the training procedure; assumed background of the SNN literature.
  • domain assumption CTC marginalization over monotonic alignments with greedy decoding gives a usable recognition objective for line-level HTR.
    Eq. (2) and decoding protocol in Appendix B.1; standard for HTR.
  • domain assumption The shared classifier head's blank posterior and entropy preview is informative enough to prune positions without removing needed separators.
    Section 3.4, Eq. (14)-(16); the auxiliary losses are designed to enforce this, but it remains a structural assumption.
  • domain assumption Handwriting priors (strokes are spatially coherent, edges supported by ink, isolated responses are noise) hold for the three Latin-script benchmarks.
    InkCoder design, Section 3.2 and Appendix A.1; acknowledged as limited for other scripts in Section E.
  • domain assumption The compute-only arithmetic proxy with fixed per-op costs and eligibility rules reflects relative energy on target event-driven hardware.
    Appendix D.3, Eq. (33)-(36); the paper itself notes it ignores memory traffic and is not a GPU/TPU energy model.

pith-pipeline@v1.3.0-daily-deepseek · 20405 in / 11268 out tokens · 106463 ms · 2026-08-04T23:36:55.011644+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Spike-HTR: Spiking Neural Transformer for Handwritten Text Recognition." pith.science (2026). https://pith.science/paper/7SK2Z5GX

@misc{pith2026260801646,
  author       = {Pith},
  title        = {Pith review of: Spike-HTR: Spiking Neural Transformer for Handwritten Text Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7SK2Z5GX}},
  note         = {Machine review of arXiv:2608.01646}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Handwritten Text Recognition (HTR) is computationally imbalanced in two ways: most image pixels are background, and many width-axis sequence positions are blank-dominated. This creates a mismatch for Spiking Neural Networks (SNNs): handwriting is observed as a static image, whereas spiking computation unfolds over timesteps. We propose Spike-HTR, a hybrid spiking recognizer that controls both the number of spiking steps and the number of width positions processed by the deep sequence mixer. To make a static image suitable for short-horizon spiking inference, InkCoder converts it into a coarse-to-fine input stream, where early steps cover broad stroke regions and later steps emphasize sharper stroke details. To reduce sequence computation, a CTC-guided length reducer keeps likely character or uncertain positions and compresses long blank-dominated stretches before deep mixing. With $T{=}2$, Spike-HTR trains only on target data, decodes without language models or lexicons, and reaches validation/test CERs of 3.5/5.4, 2.3/2.5, and 4.2/3.9 on IAM, LAM, and READ2016. Codes are available at https://github.com/QomolangmaH/SpikeHTR.

Figures

Figures reproduced from arXiv: 2608.01646 by Haoqi Zhu, Hongzhi Wang, Jinxing Han, Xiubo Liang, Yuke Li, Yu Zhao.

Figure 1
Figure 1. Figure 1: Overview of Spike-HTR architecture. A lightweight stem first extracts a shared feature map from the input line image. InkCoder derives a short sequence of spatial gates from the static image: early steps preserve broad stroke regions, and later steps emphasize sharper stroke details. A spiking 2D encoder extracts features and pools them over height to form a 1D sequence along the image width. Before the de… view at source ↗
Figure 2
Figure 2. Figure 2: Mechanism of InkCoder. A static text-line image is first converted to grayscale and decomposed into two deterministic evidence streams. The intensity stream estimates stable stroke mass through smoothing, quantile normalization, and inversion. The edge stream extracts boundary-sensitive evidence and suppresses texture-induced responses using ink-supported and spatially coherent filtering. At timestep t, th… view at source ↗
Figure 3
Figure 3. Figure 3: Temporal input streams under different static-to-temporal entry schemes (LAM sample #000883). Columns correspond to timesteps t=1, . . . , T. Gray shows the input reference; rows illustrate repeat injection, InkCoder, Poisson/rate coding, and latency coding. InkCoder produces the coarse-to-fine behavior induced by [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Event sparsity diagnostics for Spike-HTR on 512 IAM samples (d=768, T=2). Panels (a,b) show that module-wise firing rates concentrate at a few percent with a bounded tail, indicating globally sparse and well-controlled activity. Panel (c) localizes the highest firing intensities to a small subset of 1D mixer outputs. Crucially, panel (d) separates intensity from volume: despite higher firing rates in the 1… view at source ↗
Figure 5
Figure 5. Figure 5: InkCoder evidence proxies and suppression. Top row shows Xgray, the intensity proxy Iint, the normalized edge magnitude, and the edge proxy Iedge after cross-modal and coherence suppression. Bottom row illustrates the density keep mask on fused evidence and the step-conditioned evidence Dt at early and late steps. t=0 θ=0.00 t=1 θ=0.33 t=2 θ=0.67 t=3 θ=1.00 0.0 0.5 1.0 −0.5 0.0 0.5 Gate ΔGate (a) Sample LA… view at source ↗
Figure 6
Figure 6. Figure 6: InkCoder temporal dynamics. Columns correspond to steps t = 1, . . . , T with thresholds θt shown in the headers. The top row visualizes a gated-input view for intuition, X˜t = Xgray ⊙ [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Idealized compute-only energy ratio under MAC–AC cost sweeps. arithmetic becomes event-driven while the stem remains dense. Let r denote an element-weighted global average firing rate measured from recorded spikes. Then N ideal MAC = N stem MAC, Nideal AC = r [PITH_FULL_IMAGE:figures/full_fig_p019_7.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

60 extracted references · 42 canonical work pages · 1 internal anchor

  1. [1]

    International journal on document analysis and recognition , volume=

    The IAM-database: an English sentence database for offline handwriting recognition , author=. International journal on document analysis and recognition , volume=. 2002 , publisher=

    work page 2002

  2. [2]

    2022 26th International Conference on Pattern Recognition (ICPR) , pages=

    The LAM dataset: a novel benchmark for line-level handwritten text recognition , author=. 2022 26th International Conference on Pattern Recognition (ICPR) , pages=. 2022 , organization=

    work page 2022

  3. [3]

    2016 15th International conference on frontiers in handwriting recognition (ICFHR) , pages=

    ICFHR2016 competition on handwritten text recognition on the READ dataset , author=. 2016 15th International conference on frontiers in handwriting recognition (ICFHR) , pages=. 2016 , organization=

    work page 2016

  4. [4]

    2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=

    Are multidimensional recurrent layers really necessary for handwritten text recognition? , author=. 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=. 2017 , organization=

    work page 2017

  5. [5]

    2019 International conference on document analysis and recognition (ICDAR) , pages=

    A scalable handwritten text recognition system , author=. 2019 International conference on document analysis and recognition (ICDAR) , pages=. 2019 , organization=

    work page 2019

  6. [6]

    2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=

    Evaluating sequence-to-sequence models for handwritten text recognition , author=. 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=. 2019 , organization=

    work page 2019

  7. [7]

    2020 17th International Conference on Frontiers in Handwriting Recognition (ICFHR) , pages=

    Recurrence-free unconstrained handwritten text recognition using gated fully convolutional network , author=. 2020 17th International Conference on Frontiers in Handwriting Recognition (ICFHR) , pages=. 2020 , organization=

    work page 2020

  8. [8]

    International conference on document analysis and recognition , pages=

    Transformer for handwritten text recognition using bidirectional post-decoding , author=. International conference on document analysis and recognition , pages=. 2021 , organization=

    work page 2021

  9. [9]

    Pattern Recognition , volume=

    Pay attention to what you read: non-recurrent handwritten text-line recognition , author=. Pattern Recognition , volume=. 2022 , publisher=

    work page 2022

  10. [10]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Trocr: Transformer-based optical character recognition with pre-trained models , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  11. [11]

    Pattern Recognition , volume=

    HTR-VT: Handwritten text recognition with vision transformer , author=. Pattern Recognition , volume=. 2025 , publisher=

    work page 2025

  12. [12]

    Pattern Recognition , volume=

    A set of benchmarks for handwritten text recognition on historical documents , author=. Pattern Recognition , volume=. 2019 , publisher=

    work page 2019

  13. [13]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2016 , publisher=

    work page 2016

  14. [14]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  15. [15]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    Pith/arXiv arXiv 2010

  16. [16]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Swin transformer: Hierarchical vision transformer using shifted windows , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  17. [17]

    Neural networks , volume=

    Networks of spiking neurons: the third generation of neural network models , author=. Neural networks , volume=. 1997 , publisher=

    work page 1997

  18. [18]

    Neural networks , volume=

    Deep learning in spiking neural networks , author=. Neural networks , volume=. 2019 , publisher=

    2019

  19. [19]

    ACM Computing Surveys , volume=

    Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware , author=. ACM Computing Surveys , volume=. 2023 , publisher=

    work page 2023

  20. [20]

    Science , volume=

    A million spiking-neuron integrated circuit with a scalable communication network and interface , author=. Science , volume=. 2014 , publisher=

    work page 2014

  21. [21]

    Ieee Micro , volume=

    Loihi: A neuromorphic manycore processor with on-chip learning , author=. Ieee Micro , volume=. 2018 , publisher=

    work page 2018

  22. [22]

    2021 IEEE Workshop on Signal Processing Systems (SiPS) , pages=

    Efficient neuromorphic signal processing with loihi 2 , author=. 2021 IEEE Workshop on Signal Processing Systems (SiPS) , pages=. 2021 , organization=

    work page 2021

  23. [23]

    Proceedings of the IEEE , volume=

    The spinnaker project , author=. Proceedings of the IEEE , volume=. 2014 , publisher=

    work page 2014

  24. [24]

    Neural Processing Letters , volume=

    A survey of encoding techniques for signal processing in spiking neural networks , author=. Neural Processing Letters , volume=. 2021 , publisher=

    work page 2021

  25. [25]

    2021 International Joint Conference on Neural Networks (IJCNN) , pages=

    Temporal pattern coding in deep spiking neural networks , author=. 2021 International Joint Conference on Neural Networks (IJCNN) , pages=. 2021 , organization=

    work page 2021

  26. [26]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Temporal-coded deep spiking neural network with easy training and robust performance , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  27. [27]

    arXiv preprint arXiv:2512.01687 , year=

    Revisiting Direct Encoding: Learnable Temporal Dynamics for Static Image Spiking Neural Networks , author=. arXiv preprint arXiv:2512.01687 , year=

    work page arXiv

  28. [28]

    arXiv preprint arXiv:2505.18608 , year=

    Spiking Transformers Need High Frequency Information , author=. arXiv preprint arXiv:2505.18608 , year=

    arXiv

  29. [29]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Direct training for spiking neural networks: Faster, larger, better , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  30. [30]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Going deeper with directly-trained larger spiking neural networks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  31. [31]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Incorporating learnable membrane time constant to enhance learning of spiking neural networks , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  32. [32]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Towards memory-and time-efficient backpropagation for training spiking neural networks , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  33. [33]

    Advances in neural information processing systems , volume=

    Slayer: Spike layer error reassignment in time , author=. Advances in neural information processing systems , volume=

    work page

  34. [34]

    Neural computation , volume=

    Superspike: Supervised learning in multilayer spiking neural networks , author=. Neural computation , volume=. 2018 , publisher=

    2018

  35. [35]

    Advances in neural information processing systems , volume=

    Long short-term memory and learning-to-learn in networks of spiking neurons , author=. Advances in neural information processing systems , volume=

    work page

  36. [36]

    arXiv preprint arXiv:2209.15425 , year=

    Spikformer: When spiking neural network meets transformer , author=. arXiv preprint arXiv:2209.15425 , year=

    Pith/arXiv arXiv

  37. [37]

    Advances in neural information processing systems , volume=

    Spike-driven transformer , author=. Advances in neural information processing systems , volume=

    work page

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Qkformer: Hierarchical spiking transformer using qk attention , author=. Advances in Neural Information Processing Systems , volume=

    work page

  39. [39]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Spikingresformer: Bridging resnet and vision transformer in spiking neural networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  40. [40]

    European conference on computer vision , pages=

    Scene text recognition with permuted autoregressive sequence models , author=. European conference on computer vision , pages=. 2022 , organization=

    work page 2022

  41. [41]

    IEEE Signal Processing Magazine , volume=

    Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks , author=. IEEE Signal Processing Magazine , volume=. 2019 , publisher=

    2019

  42. [42]

    Spiking Neural Networks Need High-Frequency Information , author=

    work page

  43. [43]

    Advances in neural information processing systems , volume=

    Dynamicvit: Efficient vision transformers with dynamic token sparsification , author=. Advances in neural information processing systems , volume=

  44. [44]

    arXiv preprint arXiv:2202.07800 , year=

    Not all patches are what you need: Expediting vision transformers via token reorganizations , author=. arXiv preprint arXiv:2202.07800 , year=

    Pith/arXiv arXiv

  45. [45]

    ICLR , year=

    Token Merging: Your ViT But Faster , author=. ICLR , year=

    work page

  46. [46]

    International Conference on Document Analysis and Recognition , pages=

    Accelerating Transformer-Based Scene Text Detection and Recognition via Token Pruning , author=. International Conference on Document Analysis and Recognition , pages=. 2023 , organization=

    work page 2023

  47. [47]

    CTC Blank Triggered Dynamic Layer-Skipping for Efficient CTC-based Speech Recognition

    CTC Blank Triggered Dynamic Layer-skipping for Efficient CTC-based Speech Recognition , author =. arXiv preprint arXiv:2401.02046 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=

    Gated convolutional recurrent neural networks for multilingual handwriting recognition , author=. 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=. 2017 , organization=

    work page 2017

  49. [49]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    OrigamiNet: weakly-supervised, segmentation-free, one-step, full page text recognition by learning to unfold , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  50. [50]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    End-to-end handwritten paragraph text recognition using a vertical attention network , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2022 , publisher=

    work page 2022

  51. [51]

    International Conference on Document Analysis and Recognition , pages=

    Faster dan: Multi-target queries with document positional encoding for end-to-end handwritten document recognition , author=. International Conference on Document Analysis and Recognition , pages=. 2023 , organization=

    work page 2023

  52. [52]

    International Conference on Document Analysis and Recognition , pages=

    SPAN: a simple predict & align network for handwritten paragraph recognition , author=. International Conference on Document Analysis and Recognition , pages=. 2021 , organization=

    work page 2021

  53. [53]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  54. [54]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  55. [55]

    M. J. Kearns , title =

  56. [56]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  57. [57]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  58. [58]

    Suppressed for Anonymity , author=

  59. [59]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  60. [60]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959