Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Improved Robustness in AI-Generated Music Detection

T0 review · 3 major / 5 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read A log-frequency cross-correlation detector stays accurate on AI music even after speed changes, and recovers the speed factor used.

desk verdict Solid, lightweight fix for the speed-attack hole in artifact-based AI-music detectors; the DSP story is clean and the tables support it, though "invariance by construction" is slightly oversold once real resampling and band edges appear. read the letter →

arxiv 2607.27454 v1 pith:3GESTDF3 submitted 2026-07-29 cs.SD

Emile Dugelay , Thomas Barand , Baptiste Campeas , Aurélien Laouar , Darius Afchar , Romain Hennequin This is my paper
classification cs.SD
keywords AI-generatedmusicdetectionspectralartifactslog-frequencyequivariancespeedmodificationrobustnesscross-correlationfilterfakeprintinterpretable
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AI music generators leave architecture-specific energy peaks in the spectrum. Detectors that hunt those peaks work almost perfectly on raw tracks but fail when someone simply speeds the audio up or down, because the peaks move. This paper reframes that failure as a translation problem: on a log-frequency axis, a speed change becomes a rigid shift. It builds a tiny pipeline that is equivariant under that shift at the feature stage and invariant under it at the decision stage, so robustness is guaranteed by design rather than by training on every possible speed. The same lag that makes the decision also estimates how much the track was sped, giving a forensic readout. On several commercial generators the method matches prior artifact detectors on clean audio and keeps near-perfect scores under continuous speed attacks where baselines collapse.

What carries the argument

The frequency-scaling-invariant pipeline: log-STFT fakeprint extractor (equivariant) paired with one learned 1D cross-correlation filter and max-pooling (invariant), trained with a hybrid BCE-plus-lag CE loss.

What would settle it

Take Suno or Udio tracks, apply continuous real-world speed changes (including ordinary resampling pipelines), recompute the log-fakeprint, and check whether the cross-correlation peak still lines up with the true speed factor and whether detection F1 stays near the reported 0.986; systematic misalignment or collapse would refute the claim.

Watch

Extended reading notes

Core claim

Mapping the time-averaged STFT onto a log-frequency axis turns speed modification into a pure translation of the artifact pattern; a single learned cross-correlation filter plus max-pooling then makes the binary decision invariant to that translation, so the detector remains accurate under speed change by construction and the argmax lag recovers the applied speed factor.

Load-bearing premise

Real speed changes must move the isolated artifact peaks almost exactly like an ideal frequency scaling on the log axis, without anti-alias filters or decoding burying or warping them.

Editorial extensions

If this is right

  • Streaming platforms can screen for AI music after common speed edits without retraining on every speed factor.
  • The lag readout supplies a forensic estimate of how much a track was sped, useful when contesting false positives.
  • The same construction gives partial robustness to pitch shift, because pitch shift ends with a resampling step that the log-axis already absorbs.
  • Because the model has only a few thousand parameters and trains in minutes, covering a new generator architecture is cheap.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Any other post-process that multiplies all frequencies by a constant (not only artistic speed-up) should be handled the same way without new labels.
  • Non-uniform spectral warps such as EQ or multi-band compression would still break the rigid-shift premise and likely need a different equivariance.
  • If future generators remove transposed-convolution upsampling, the fakeprint itself may vanish and the method would need a new artifact source.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a lightweight, interpretable detector for AI-generated music that is intended to be robust to speed modification by construction. Building on Afchar et al.’s spectral-artifact (fakeprint) analysis, it remaps a time-averaged STFT onto a log-frequency axis so that ideal frequency scaling becomes a translation, then classifies with a single learned cross-correlation filter plus max-pooling (shift invariance) and a hybrid BCE+CE loss that also localizes the lag. Experiments on Suno v3.5/v5 and Udio v120 show near-SOTA clean performance and strong robustness under continuous speed factors in [0.7, 1.4], including a Clean-train/Attack-test cell with high AUC; the argmax lag recovers α with very low MAE. Partial robustness to pitch shift is also reported.

Significance. If the result holds, this is a clear, useful advance on a documented failure mode of both content-based and artifact-based AI-music detectors. The contribution is attractive on several axes the community values: (i) an explicit equivariance/invariance formulation rather than pure augmentation, (ii) a few-thousand-parameter model with released code and sub-5-minute training, (iii) dual output (binary decision + estimated speed factor) that supports forensic use, and (iv) systematic Clean×Attack tables across three generators plus λ ablation and out-of-range α curves. Even with caveats on ideal scaling, the empirical robustness gap versus Afchar et al. and SpecTTTra-α is large enough to matter for platform monitoring.

major comments (3)
  1. [§3.4, §4.1–4.2, Table 2 Clean/Attack] §3.4 Eq. (3)–(4) and §4.1–4.2 frame score-level robustness as holding by construction once A is log-frequency equivariant and D max-pools the cross-correlation. Table 2 Clean/Attack contradicts pure score invariance: on Suno v5, AUC stays 0.986 but F1 falls from 0.998 to 0.732 (precision 0.993 ⇒ recall ≈0.58) at the trained bias b. If x were a rigid circular translate, max_k z_k + b would be unchanged and the operating point would transfer. The drop implies real resampling, fixed analysis bands (peaks leaving 5–16 kHz / 1–7 kHz), log-interpolation, or non-equivariant lower-hull on a bounded axis distort peak heights. Attack/Attack near-perfect F1 therefore still depends on bin-aligned speed augmentation and the CE term (λ=0.05), which the paper contrasts with “augmentation races” (§2, §6.3). Please either (a) empirically measure how well A∘T_α ≈ translate(A) (e.g., correlation / peak-hei
  2. [§3.3–3.4, §4.1, §6.2 lag MAE] The ideal map T_α:s(t)↦s(αt) in §3.3 is carried unchanged into the extractor design, but training/test speed changes are implemented as practical resampling. Anti-alias filtering, generator decode chains, and content-dependent lower-hull subtraction can break rigid translation of P on the log axis. Without a controlled check (synthetic Dirac-comb or known-architecture peaks under the same resampler used at test), both the invariance claim and the lag-as-α readout (MAE 4e-4) rest on an untested premise. A short diagnostic—peak location error and max(z) stability vs α on held-out clean fakes—would load-bear the central story.
  3. [Abstract, §1, Table 3, §6.2–6.3] Table 3 (pitch shift) and §6.3 correctly note limited robustness, but the abstract and intro still lead with “prevent this kind of attack by design” alongside pitch shifting. Pitch shift is time-stretch + resample; only the resample part is addressed. Either narrow the abstract claim to speed/resampling or quantify how much of the Table 3 degradation is stretch vs scale, and whether threshold recalibration (as suggested in §6.2) restores usable F1 without retraining.
minor comments (5)
  1. [Table 1, §6.1] Table 1 only sweeps λ∈{0,0.05,0.10} on Attack/Attack Suno v5; a brief note on whether λ=0.05 remains best under Clean/Attack and on other generators would strengthen the “report only λ=0.05” choice.
  2. [Figure 2] Figure 2 is helpful; adding the numerical k* / α̂ on the bottom panel would make the forensic readout immediately readable.
  3. [§5.2] §5.2: native-rate handling (16 kHz SONICS vs 44.1 kHz Suno v5) and different analysis bands are reasonable but should state whether any cross-rate evaluation was tried, or that models are rate-specific.
  4. [§1, §5.2, References] Typos/clarity: “Producer” in §1 likely needs a product name or citation; “B'” vs “1920 bins per octave” relationship could be stated once in closed form; references [1] and Deezer newsroom dated 2026 may need stable archival links for the camera-ready.
  5. [Table 2, §6.2] SpecTTTra-α is evaluated as released and is weak on Suno v5 clean; a one-sentence caveat that it was not retrained on v5 (already footnoted) could also appear in the main Clean/Clean discussion so readers do not over-read the gap.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: log-frequency equivariance and max-pool shift invariance are standard constructions; empirical robustness is measured on held-out continuous-α attacks, including zero-shot Clean/Attack.

full rationale

The paper's central derivation is architectural, not definitional or fitted-as-prediction. Speed change T_α scales frequencies by 1/α; on ξ = log f this is a translation (Eq. 3). The extractor (STFT → time average → log remap → lower-hull) is built to realize that equivariance under ideal scaling; the detector (single cross-correlation filter + max over lags + bias) is translation-invariant by the max, so D∘A is score-invariant when the premise holds. That is ordinary change-of-variables plus max-pooling, not X defined as Y. The learned filter w and bias b are fit from data; detection AUC/F1 and lag MAE are then reported on held-out tracks with continuous (non-bin-aligned) speed factors, including a Clean-train/Attack-test cell never optimized for attack. The auxiliary CE term supervises bin-aligned lag only on AI training samples and is ablated (λ); it regularizes boundary weights and slightly improves metrics but does not force the headline Attack/Attack or Clean/Attack numbers by construction. Citation of Afchar et al. [2] (overlapping authors) supplies the prior artifact characterization that motivates the fakeprint; it is not a uniqueness theorem that forbids alternatives, and the present contribution—log-axis equivariance + cross-correlation invariance and the speed-factor readout—is independently specified and benchmarked against external baselines (Afchar linear-axis detector, SpecTTTra). Clean/Attack F1 drop at fixed bias is an empirical limit of real resampling/band truncation versus ideal T_α, i.e. a correctness/robustness gap, not a circular reduction of claim to input. No self-definitional loop, no fitted quantity renamed as prediction, no load-bearing uniqueness import. Score 0; steps empty.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central robustness claim rests on standard DSP identities (scaling↔log translation), on the prior architectural-fingerprint theory from Afchar et al., and on modeling choices (lower-hull isolation, single shared template w, discrete bin-aligned α for CE targets). Free parameters are ordinary training knobs (λ, binning, bands, bias b) rather than physics-style fitted constants. No new physical entities are postulated; ‘fakeprint’ is a named representation of already-described peaks.

free parameters (5)
  • λ (CE lag-loss weight) = 0.05
    Balances BCE detection vs CE peak localization; chosen by small grid (0, 0.05, 0.10) on attacked Suno v5; λ=0.05 reported thereafter.
  • log-frequency resolution B' / F bins (1920 bins per octave) = 1920 bins/octave
    Sets discrete shift granularity and CE target indices; chosen as engineering resolution, not derived.
  • analysis band limits (1–7 kHz SONICS; 5–16 kHz Suno v5) = 1–7 kHz or 5–16 kHz by source
    Inherited/adapted from prior fingerprint work and sampling-rate constraints; directly affects which peaks are visible.
  • N_fft = 2^14 and trainable bias b = N_fft=16384; b learned
    STFT size and decision threshold offset are standard hyperparameters of the pipeline.
  • speed-factor training support [~0.7, ~1.4] / ±990 bins = ×0.7 to ×1.4 (bin-aligned train)
    Defines the Attack training distribution and CE supervision range; evaluation partly outside but still nearby.
assumptions (5)
  • domain assumption Transposed-convolution upsampling in neural audio decoders produces architecture-determined spectral peak combs (fingerprints) independent of weights/data.
    Taken from Afchar et al. [2] and used as the reason a template w exists (§3.1, §2).
  • standard math Under ideal speed change T_α, the log-frequency fakeprint translates rigidly by −log α (Eq. 3).
    Standard change-of-variable on ξ=log f; load-bearing for equivariance-by-construction (§3.4).
  • domain assumption Lower-hull / piecewise-linear baseline subtraction isolates artifact peaks from musical content sufficiently for detection.
    Carried from [2] into Eq. 6; if content leaks into A(s), both D and α estimates degrade.
  • ad hoc to paper A single shared 1D template w plus max_k z_k is a sufficient statistic for binary AI vs human on these generators.
    Architectural bet of §4.2; justified empirically on three generators, not proved generally.
  • ad hoc to paper Bin-aligned discrete α in training yields an unambiguous CE target that still generalizes to continuous α at test.
    Stated in §5.1; underpins the lag-supervision design.
invented entities (1)
  • fakeprint (log-frequency isolated artifact representation A(s)) independent evidence
    purpose: Name the 1D log-frequency peak map that the cross-correlation template matches.
    Terminological packaging of the Afchar-style residual spectrum after log remap; not a new physical object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improved Robustness in AI-Generated Music Detection." pith.science (2026). https://pith.science/paper/3GESTDF3

@misc{pith2026260727454,
  author       = {Pith},
  title        = {Pith review of: Improved Robustness in AI-Generated Music Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GESTDF3}},
  note         = {Machine review of arXiv:2607.27454}
}
read the original abstract

AI music generators leave predictable spectral artifacts determined by their architecture. Existing detectors exploit these artifacts with near-perfect accuracy on raw generated tracks, but their performance collapses under simple audio manipulations, such as speed modification or pitch shifting. We address this open robustness problem by introducing a frequency-scaling-invariant detection pipeline that aims to prevent this kind of attack by design. Our method maps audio onto a log-frequency axis via a log-STFT remapping. A single learned cross-correlation filter, combined with max-pooling, provides shift invariance at inference time. Training uses a hybrid loss that jointly supervises binary detection and artifact-peak localization, regularizing boundary weights. Because robustness to speed change is built in by design, the detector is also interpretable: it outputs both a binary decision and an estimate of the applied speed-change factor.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 2 linked inside Pith

  1. [1]

    Improved robustness in AI-generated music detection

    INTRODUCTION Audio GenAI tools like Suno, Udio, or Producer empower users with easy, prompt-based song creation. Although these tools open up creative possibilities, AI-generated tracks are widely used to flood streaming platforms as well as to drive fraudulent traffic as a malicious way to divert revenue from real artists [1]. In the absence of transpar-...

    2026

  2. [2]

    upsampling artifacts

    RELATED WORK The detection of AI-generated music is an emerging MIR task sharing methodological overlap with composer iden- tification and music auto-tagging [3], as well as deepfake detection [7–9]. Existing approaches can be categorized into two families. Content-based detection.The first family relies on learned representations of the musical content, ...

    arXiv 2026

  3. [3]

    As shown in [2], this upsampling process causes the signal spectrum to alias, producing spurious energy peaks at pre- dictable frequencies

    PROBLEM FORMULATION 3.1 Spectral Artifacts in AI-Generated Music Most AI music generators rely on transposed convolution layers to upsample latent embeddings into waveforms. As shown in [2], this upsampling process causes the signal spectrum to alias, producing spurious energy peaks at pre- dictable frequencies. For a single deconvolution layer of stridek...

    2026

  4. [4]

    PROPOSED METHOD We now describe the concrete implementation of the two- stage pipeline introduced in Section 3. The artifact extrac- torAis engineered to satisfy log-frequency equivariance, the detectorDis built to be invariant to such translations, so that the full pipelineD◦Aremains robust to time- frequency scalingT α. An overview of both components is...

    2026

  5. [5]

    Clean" Dataset:Consists of the original audio tracks without any modifications. •

    EXPERIMENTAL SETUP 5.1 Dataset To train our network to distinguish between AI-generated and genuine audio, we compiled a dataset of 10,000 audio tracks. Each track was truncated to retain only the first 30 seconds of audio. The dataset is evenly balanced, 5,000 AI-generated tracks sourced from Suno v5 and 5,000 gen- uine tracks sampled from the FMA-small ...

    2026

  6. [6]

    [2] and the SpecTTTra-α(best performing model of [10]),(2)Robust- ness to speed change attacks across all model variants

    RESULTS & DISCUSSION We evaluate our method across two main axes:(1)De- tection performance on unattacked audio compared to two baselines: our implementation of Afchar et al. [2] and the SpecTTTra-α(best performing model of [10]),(2)Robust- ness to speed change attacks across all model variants. Finally, we show that the robustness to speed modifica- tion...

    arXiv 2026

  7. [7]

    CONCLUSION This paper addresses a critical vulnerability in current AI music detectors: their drastic accuracy drop under sim- ple frequency-scaling manipulations such as speed mod- ification. We introduced a mathematically grounded, lightweight pipeline that is frequency-scaling-invariant by design: mapping audio to a log-frequency axis yields shift equi...

    2026

  8. [8]

    Reliable detection contributes to a fairer distribution of revenue toward human artists and to greater transparency for listeners

    ETHICS STATEMENT The detector proposed in this paper is designed to sup- port content monitoring on streaming platforms, where AI- generated tracks are increasingly used to extract fraudulent royalties at scale. Reliable detection contributes to a fairer distribution of revenue toward human artists and to greater transparency for listeners. Despite these ...

Show all 32 references
  1. [9]

    Deezer confirms demonetization of up to 85% of AI-music streams due to fraud and moves to sell AI-detection Tech- nology,

    Deezer, “Deezer confirms demonetization of up to 85% of AI-music streams due to fraud and moves to sell AI-detection Tech- nology,” https://newsroom-deezer.com/2026/01/ ai-generated-music-deezer-selling-detection-tool/, 2026, [Online; accessed 24-April-2026]

    2026

  2. [10]

    A Fourier Explanation of AI-music Ar- tifacts,

    D. Afchar, G. Meseguer-Brocal, K. Akesbi, and R. Hennequin, “A Fourier Explanation of AI-music Ar- tifacts,” inISMIR, 2025

    2025

  3. [11]

    The AI Music Arms Race: On the Detection of AI- Generated Music,

    L. C. Vila, B. L. T. Sturm, L. Casini, and D. Dalmazzo, “The AI Music Arms Race: On the Detection of AI- Generated Music,”Transactions of the International Society for Music Information Retrieval, vol. 8, 2025

    2025

  4. [12]

    AI-Generated Music Detection and its Challenges,

    D. Afchar, G. Meseguer-Brocal, and R. Hennequin, “AI-Generated Music Detection and its Challenges,” Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  5. [13]

    Sped-up y nightcore en el paradigma de las plataformas,

    E. Viñuela, “Sped-up y nightcore en el paradigma de las plataformas,”Revista de la Asociación Española de Investigación de la Comunicación, vol. 12, no. 24, p. raeic122404, Nov. 2025. [Online]. Available: https://revistaeic.eu/index.php/raeic/article/view/686

    2025

  6. [14]

    TikTokers are speeding up songs - but are we forgetting the originals?

    Christian Brooks BBC News, “TikTokers are speeding up songs - but are we forgetting the originals?” https://www.bbc.com/news/articles/cqv5x2qe8q6o, 2024, [Online; accessed 24-April-2026]

    2024

  7. [15]

    ASVspoof: the automatic speaker verification spoofing and countermeasures challenge,

    Z. Wu, J. Yamagishi, T. Kinnunen, C. Hanilçi, M. Sahidullah, A. Sizov, N. Evans, M. Todisco, and H. Delgado, “ASVspoof: the automatic speaker verification spoofing and countermeasures challenge,” IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 4, pp. 588–604, 2017

    2017

  8. [16]

    The creation and detection of deepfakes: A survey,

    Y . Mirsky and W. Lee, “The creation and detection of deepfakes: A survey,”ACM computing surveys (CSUR), vol. 54, no. 1, pp. 1–41, 2021

    2021

  9. [17]

    CNN-generated images are surprisingly easy to spot... for now,

    S.-Y . Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, “CNN-generated images are surprisingly easy to spot... for now,” inProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2020, pp. 8695–8704

    2020

  10. [18]

    SONICS: Synthetic or not – iden- tifying counterfeit songs,

    M. A. Rahman, Z. I. A. Hakim, N. H. Sarker, B. Paul, and S. A. Fattah, “SONICS: Synthetic or not – iden- tifying counterfeit songs,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  11. [19]

    Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,” inICASSP. IEEE, 2023, pp. 1–5

    2023

  12. [20]

    Fake- musiccaps: A dataset for detection and attribution of synthetic music generated via text-to-music models,

    L. Comanducci, P. Bestagini, and S. Tubaro, “Fake- musiccaps: A dataset for detection and attribution of synthetic music generated via text-to-music models,” Journal of Imaging, vol. 11, no. 7, p. 242, 2025

    2025

  13. [21]

    Singfake: Singing voice deepfake detection,

    Y . Zang, Y . Zhang, M. Heydari, and Z. Duan, “Singfake: Singing voice deepfake detection,” in ICASSP. IEEE, 2024

    2024

  14. [22]

    From Audio Deepfake Detection to AI-Generated Music Detection–A Pathway and Overview,

    Y . Li, M. Milling, L. Specia, and B. W. Schuller, “From Audio Deepfake Detection to AI-Generated Music Detection–A Pathway and Overview,”arXiv preprint arXiv:2412.00571, 2024

    2024 arXiv

  15. [23]

    Double entendre: Robust audio-based AI-generated lyrics detection via multi-view fusion,

    M. Frohmann, G. Meseguer-Brocal, M. Schedl, and E. V . Epure, “Double entendre: Robust audio-based AI-generated lyrics detection via multi-view fusion,” inFindings of the Association for Computational Lin- guistics: ACL 2025, 2025, pp. 1914–1926

    2025

  16. [24]

    AI-Generated Music Detection in Broadcast Monitoring,

    D. Lopez-Ayala, A. Cabello, P. Zinemanas, E. Molina, and M. Rocamora, “AI-Generated Music Detection in Broadcast Monitoring,” inICASSP, 2026

    2026

  17. [25]

    Segment Transformer: AI- Generated Music Detection via Music Structural Anal- ysis,

    Y . Kim and S. Go, “Segment Transformer: AI- Generated Music Detection via Music Structural Anal- ysis,” in2025 Asia Pacific Signal and Information Pro- cessing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2025, pp. 664–669

    2025

  18. [26]

    De- convolution and checkerboard artifacts,

    A. Odena, V . Dumoulin, and C. Olah, “De- convolution and checkerboard artifacts,”Distill,

  19. [28]

    Adver- sarial Audio Synthesis,

    C. Donahue, J. McAuley, and M. Puckette, “Adver- sarial Audio Synthesis,”International Conference on Learning Representations (ICLR), 2019. 7 Proceedings of the 27th ISMIR Conference, Abu Dhabi, UAE, November 08–12, 2026

    2019

  20. [29]

    Upsam- pling Artifacts in Neural Audio Synthesis,

    J. Pons, S. Pascual, G. Cengarle, and J. Serrà, “Upsam- pling Artifacts in Neural Audio Synthesis,” inICASSP, 2021

    2021

  21. [30]

    Melgan: Generative adversarial net- works for conditional waveform synthesis,

    K. Kumar, R. Kumar, T. de Boissiere, L. Gestin, W. Z. Teoh, J. Sotelo, A. de Brebisson, Y . Bengio, and A. Courville, “Melgan: Generative adversarial net- works for conditional waveform synthesis,”Advances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  22. [31]

    FMA: A dataset for music analysis,

    K. Benzi, M. Defferrard, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” CoRR, 2016

    2016

  23. [32]

    Rubber band library: Au- dio time stretcher library,

    Particular Programs Ltd, “Rubber band library: Au- dio time stretcher library,” https://breakfastquay.com/ rubberband/, 2024, version 4.0.0. 8

    2024

  24. [2016]

    Available: http://distill.pub/2016/ deconv-checkerboard

    [Online]. Available: http://distill.pub/2016/ deconv-checkerboard

    2016

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.