Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Post-hoc latent restructuring can tailor frozen audio autoencoders without retraining them.

desk verdict Honest, useful empirical method paper; the latent-only transfer claim needs a second frozen checkpoint and error bars before I'd trust it as a general recipe. read the letter →

arxiv 2507.07867 v2 pith:H4CBA53B submitted 2025-07-10 cs.SD cs.LGeess.AS

Dimitrios Bralios , Jonah Casebeer , Paris Smaragdis This is my paper · ORCID
classification cs.SDcs.LGeess.AS
keywords audioautoencoderslatentspacerestructuringpost-hocadaptationlatent-spacelossesorderedrepresentationssemanticalignmentequivariancediffusion
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

The paper sets out to show that the latent space of an already-trained neural audio autoencoder can be reorganized after the fact by a small inner autoencoder trained entirely on latent-space losses. This matters because audio codecs are usually trained only for reconstruction, leaving their latent codes poorly matched to tasks such as generation, semantic retrieval, or source separation, and retraining an entire codec is prohibitively expensive. The authors demonstrate three restructurings: an ordered hierarchy of channels, alignment of latents to semantic embeddings, and equivariance to low-pass filtering. If the demonstration holds, any existing audio codec can be adapted to a new task with under a third of a percent of its original training compute.

What carries the argument

The carrying object is the Re-Bottleneck, an inner autoencoder (encoder $R_E$, decoder $R_D$) placed inside the frozen base autoencoder's VAE bottleneck, together with a latent discriminator $D$. The key trick is that every loss in Eq. (7) acts on the base latent $z = A_E(x)$ or the inner latent $\tilde{z}=R_E(z)$: reconstruction, adversarial and feature-matching terms keep the reconstructed $\hat{z}=R(z)$ close to real base latents, a KL term regularizes $\tilde{z}$, and the task loss $L^R_{\text{task}}$ directly sculpts $\tilde{z}$ (nested dropout in Eq. (8), contrastive InfoNCE in Eq. (9), or the equivariance loss in Eq. (10)). Because the base decoder is frozen, all structural change is forced into the learned inner latent, preserving the original decoder's waveform quality while changing what the latent code means.

What would settle it

Hold out a test set and train Re-Bottlenecks with increasing task-loss weight ($\lambda_{\text{task}} = 2.5$, $5$, $10$) while keeping the other weights fixed; the central claim would fail if latent-space reconstruction losses stay low but decoded-audio STFT/mel distance and SISDR degrade sharply, since that would show the frozen decoder amplifies latent shifts. A second check: use the equivariant variant at cutoff frequencies outside the trained 1.4-5.5 kHz range and compare the decoded output's spectral envelope to the low-pass-filtered target; large divergence would reveal that $h_\alpha$ and the learned commutation only hold inside the training distribution.

Watch

Extended reading notes

Core claim

The central claim is that a frozen audio autoencoder can be given a new bottleneck representation without modifying or retraining the base model. The Re-Bottleneck is a 19.1M-parameter inner autoencoder inserted between the base encoder and decoder: its own encoder maps the base latent $z$ to a structured inner latent $\tilde{z}$, and its own decoder reconstructs $z$ from $\tilde{z}$, with a latent discriminator keeping the reconstructed latents in-distribution. All training objectives, including the task-specific structural losses, are computed on latent vectors or discriminator features; no waveform, STFT, or perceptual audio loss is used. The paper reports three demonstrations: nested-dropout training produces channels ordered by importance with low cross-channel correlation; a contrastive loss against a semantic embedding target reaches 0.70 CKA/0.78 PWCCA, nearly matching the 0.69/0.83 semantic-only upper bound while increasing reconstruction error by roughly 5%; and a parametric latent filter trained with an explicit equivariance loss decodes close to the low-pass-filtered target (0.98 vs 0.88 STFT distance at a 1.4 kHz cutoff), while direct filtering of the base latent fails badly (5.08). This is summarized as a general post-hoc path to structured audio latents that costs under 48 GPU hours per variant, under 0.33% of the base autoencoder's training compute.

Load-bearing premise

The framework assumes that training only on latent-space signals, with the base audio decoder frozen, is enough to guarantee that the restructured latents decode into faithful audio; if the frozen decoder amplifies small latent distribution shifts, the method's central promise fails.

Editorial extensions

If this is right

  • Ordered Re-Bottlenecks give a learned nonlinear channel ordering that beats random dropout and PCA when channels are progressively removed, with a diagonal channel-correlation matrix indicating decorrelation.
  • Semantically aligned latents reach nearly the semantic-only upper bound on representation alignment (0.70 vs 0.69 CKA) while keeping reconstruction loss small, and the aligned latents improve downstream diffusion, lowering FAD from 0.435 to 0.411 and raising the prompt-alignment score from 0.185 to 0.191.
  • Equivariant Re-Bottlenecks can perform Gaussian low-pass filtering directly in latent space; at a 1.4 kHz cutoff the decoded output is close to the filtered target autoencoded by the base model (0.98 vs 0.88 STFT distance), whereas filtering the base latent directly fails (5.08).
  • Each restructuring costs under 48 GPU hours, which is under 0.33% of the 14.5K GPU hours used to train the base autoencoder, making repeated latent-prototyping experiments practical.

Reading between the lines

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

  • Inference: because the task loss is a plug-in term, the same machinery should be able to impose other user-chosen structures such as sparsity, disentanglement by source, or invariance to noise, provided a differentiable latent-domain objective can be written.
  • Inference: a text-aligned variant did not improve diffusion FAD even though the audio-semantic variant did, which suggests that raw alignment strength is not sufficient; the semantic space's compatibility with the generation target and the diffusion prior also matters.
  • Inference: if the latent-only transfer assumption holds broadly, a deployed codec could be upgraded by swapping a small adapter while the core encoder and decoder stay untouched, enabling cheap A/B testing of representational changes in production systems.
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 introduces Re-Bottleneck, a post-hoc method for restructuring the latent space of a pre-trained, frozen audio autoencoder. An inner autoencoder (Re-Encoder) is trained entirely with latent-domain losses—reconstruction, adversarial, feature-matching, and KL terms—plus optional task-specific losses—to impose ordered channels, semantic alignment, or equivariance. The method is demonstrated on the Stable Audio Open (SAO) variational autoencoder. Experiments show that ordered-channel Re-Bottlenecks outperform PCA and random dropout at low channel counts, semantically aligned variants achieve CKA/PWCCA close to a semantic-only upper bound while preserving most reconstruction quality, and an equivariant variant can perform low-pass filtering in the latent space with decoded output close to the filtered target. The authors also report diffusion training experiments suggesting that a BEATs-aligned variant improves FAD and CLAP scores over the baseline.

Significance. If the results hold, Re-Bottleneck is a practically useful contribution: it offers a cheap (sub-48-GPU-hour) way to prototype structured latent spaces for off-the-shelf audio autoencoders without retraining the base model, and the paper explicitly provides code. The evaluation metrics (CKA, PWCCA, FAD, SISDR) are external to the training objective, which strengthens the empirical claims. The three experiments cover distinct structural properties and the equivariance demonstration in particular is conceptually appealing. However, the central claim—that latent-only training transfers to waveform-level behavior through a frozen decoder—rests on a single autoencoder checkpoint, and the diffusion improvement is reported without error bars. The contribution is therefore more of a promising proof-of-concept than a fully established method.

major comments (3)
  1. [Section 3.4 / Table 1] The central claim that latent-only training preserves reconstruction fidelity is supported only by a single frozen autoencoder (SAO), with no error bars and no repeated runs. Since all gradients for the Re-Encoder flow through latent-domain terms (Eqs. 3–7) and the base decoder A_D is frozen, near-full reconstruction depends on the untested assumption that A_D does not amplify small distributional shifts in z_hat. This is load-bearing because all three demonstrations are evaluated by decoding through A_D. Please add a second frozen autoencoder (e.g., DAC or EnCodec) and report means and standard deviations over at least three seeds for the Table 1 metrics, including a full-channel reconstruction number for the ordered variant.
  2. [Section 2.3.3 / Eq. (10)] The equivariance claim rests on the hand-designed latent filter h_alpha(c) = exp(-0.5 mel(c)^1.4 alpha^-1.4), but the paper neither justifies the exponent 1.4 nor the mel-frequency mapping of latent channel index c; SAO's latent channels are learned and are not frequency-ordered by construction. The ablation in Sec. 4.3 shows that the explicit loss (Eq. 10) is necessary, but it does not validate this functional form: the Re-Encoder could be learning to compensate for an incorrect h_alpha rather than learning a genuine equivariant structure. Please compare alternative parameterizations (e.g., h_alpha linear in c, a learned affine mapping, or a channel-to-frequency assignment derived by probing) and report the equivariance error, ||h_alpha(R_E(z)) - R_E(A_E(g_alpha(x)))||, for each.
  3. [Section 4.4 / Fig. 4] The diffusion results are single runs with no confidence intervals. The reported improvement of the BEATs-aligned Re-Bottleneck over the baseline (FAD 0.411 vs 0.435) is small and may lie within the run-to-run variability of 100K-step diffusion training; as written, the conclusion that semantic alignment improves downstream diffusion is not statistically supported. Please provide at least three seeds per variant with means and standard deviations (or an equivalent significance test), or temper the claim to a qualitative observation that the semantic variant remains competitive while improving representation-alignment metrics.
minor comments (5)
  1. [Title page] The running header reads 'LA TENT' instead of 'LATENT'; please correct the typo.
  2. [Section 2.3.2] The word 'constrastive' should be 'contrastive' in the sentence introducing Eq. (9).
  3. [Section 4.2] The claim of '5% degradation' is not tied to a specific reconstruction metric; Table 1 shows different relative degradations for STFT (1.6%), mel (2.6%), and SISDR (4.6%), so please clarify which metric and whether the comparison is relative or absolute.
  4. [Table 3] Please describe precisely how the 'SAO + Latent Filtering' baseline is constructed; it is currently unclear whether the same h_alpha is applied to the original SAO latents and whether any channel reordering is used.
  5. [Section 4.1] The claim that training required 'under 0.33% of the 14.5K GPU hours' compares single-GPU hours with multi-GPU hours; please report normalized GPU-hours (e.g., total GPU-hours) to make the comparison fair.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claim is tested against external waveform-domain and semantic benchmarks, and the only self-citation (the Re-Encoder architecture) is not load-bearing.

full rationale

The paper does not dress up fitted quantities as predictions. The Re-Encoder is trained with latent-only losses (Eqs. 3-7), but every headline result is evaluated with quantities that are not identical to any training loss: STFT/mel/SISDR at the waveform level (Table 1, Fig. 1), FAD/CLAP on a downstream diffusion pipeline (Fig. 4), and CKA/PWCCA against independently pre-trained semantic encoders (Table 2). The equivariance experiment deserves the closest scrutiny because Eqs. 10-11 explicitly train R_D(h_alpha(R_E(z))) to match A_E(g_alpha(x)), so Table 3's 'ReBot vs. AE Target' comparison is partly a check that the training objective converged and that the frozen decoder preserves the latent-domain match. This is a supervised demonstration, not an independent prediction, and it includes a fair control ('SAO + Latent Filtering') showing the behavior is not a property of the base VAE alone. The only self-citation is reference [17] for the Re-Encoder architecture and the general idea of latent-domain re-encoding; it is cited as inspiration, not as a uniqueness theorem or as the sole justification for the method, and the paper supplies its own external empirical validation. The hand-designed h_alpha(c) = exp(-0.5 mel(c)^1.4 alpha^-1.4) is an explicit modeling choice rather than a hidden ansatz, and the paper ablates the explicit equivariance loss. The lack of error bars and the reliance on a single frozen SAO checkpoint are correctness/robustness concerns, not evidence that any derivation is circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The framework is empirical; its central claims rest on hyperparameters chosen by hand, a frozen pretrained autoencoder, and a hand-designed latent filter. No new physical or ontological entities are introduced. The largest unproven premise is that latent-only training preserves waveform-level fidelity through the frozen decoder.

free parameters (4)
  • Loss weights lambda_rec, lambda_kl, lambda_adv, lambda_fm = 1.0, 1e-4, 0.5, 1
    Chosen by hand in Section 3.1; the paper asserts they are easy to balance but does not systematically study sensitivity.
  • InfoNCE temperature tau = 0.07
    Standard value chosen by hand; it controls the sharpness of the semantic alignment objective.
  • Nested dropout mask application rate = 75% of batch
    In the ordered variant the mask is applied to only 75% of the batch; no ablation of this fraction is reported.
  • Latent filter h_alpha exponent and mel mapping = exponent 1.4, minimum cutoff 1.4 kHz
    Hand-designed to mimic the Gaussian STFT filter G_alpha; the equivariance results depend on this mapping, yet it is not validated independently.
assumptions (4)
  • domain assumption The Stable Audio Open VAE is a high-quality frozen base autoencoder with 64 latent channels at 21.5 Hz.
    Section 3.1 uses the released SAO checkpoint as ground truth; the method's generality to other autoencoders is asserted, not demonstrated.
  • domain assumption Latent-domain losses suffice to preserve reconstruction quality through the frozen decoder.
    This is the core assumption of Section 2.2; no waveform loss is ever computed, and the assumption is only checked on SAO.
  • domain assumption BEATs and T5 embeddings capture semantic content relevant to downstream audio generation.
    Used as alignment targets in Section 2.3.2; if these embeddings are poor semantic proxies, the semantic alignment results are weaker.
  • ad hoc to paper h_alpha(c) = exp(-0.5 mel(c)^1.4 alpha^-1.4) models the effect of Gaussian low-pass filtering in the latent space.
    This hand-designed relation between latent channel index and mel frequency is central to the equivariance experiment and is never independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders." pith.science (2026). https://pith.science/paper/H4CBA53B

@misc{pith2026250707867,
  author       = {Pith},
  title        = {Pith review of: Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4CBA53B}},
  note         = {Machine review of arXiv:2507.07867}
}
read the original abstract

Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most are trained to maximize reconstruction fidelity, often neglecting the specific latent structure necessary for optimal performance in diverse downstream applications. We propose a simple, post-hoc framework to address this by modifying the bottleneck of a pre-trained autoencoder. Our method introduces a "Re-Bottleneck", an inner bottleneck trained exclusively through latent space losses to instill user-defined structure. We demonstrate the framework's effectiveness in three experiments. First, we enforce an ordering on latent channels without sacrificing reconstruction quality. Second, we align latents with semantic embeddings, analyzing the impact on downstream diffusion modeling. Third, we introduce equivariance, ensuring that a filtering operation on the input waveform directly corresponds to a specific transformation in the latent space. Ultimately, our Re-Bottleneck framework offers a flexible and efficient way to tailor representations of neural audio models, enabling them to seamlessly meet the varied demands of different applications with minimal additional training.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [1]

    INTRODUCTION Neural audio autoencoders and codecs have become foundational in modern audio processing, enabling high-fidelity reconstruction from compact latent representations [1, 2, 3, 4]. These models are pow- erful components, driving innovation in areas like audio generation via next-token prediction [5, 6, 7], latent diffusion [4] and support- ing t...

    work page Pith review arXiv 2025

  2. [2]

    real” latents (those directly fromA E(x)and “fake

    PROPOSED METHOD 2.1. Overview Our framework employs a pre-trained and frozen neural audio au- toencoder, comprising an encoderAE and decoderA D. The encoder transforms a potentially multi-channel input waveformx∈R N×L to a compact latent representationz=A E(x)∈R C×T . The decoder reconstructs the waveform from the latent representation, ˆx=A D(z). As is c...

    work page

  3. [3]

    EXPERIMENTAL SETUP To validate the efficacy of Re-Bottlenecks, we conduct three experi- ments corresponding to the three Re-Bottleneck types above. 3.1. Re-Bottleneck Experiments For all experiments, we use the publicly released Stable Audio Open (SAO) V AE [4] as our frozenA. This V AE compresses stereo 44.1KHz audio into64channel latents at21.5Hz. We us...

    work page

  4. [4]

    modern day PCA

    EXPERIMENTS & RESULTS This section demonstrates the flexibility of the Re-Bottleneck frame- work by showcasing various latent space modifications. Training each Re-Bottleneck for these demonstrations required less than 48 GPU hours, representing under 0.33% of the 14.5K GPU hours used for the base autoencoder. 4.1. Ordered Re-Bottleneck In this experiment...

    work page

  5. [5]

    Our proposed Re-Bottleneck framework offers a flexible, post-hoc solution to this limitation

    CONCLUSION While pre-trained neural audio codecs excel at reconstruction, their default latent spaces often lack the specific structure needed for op- timal performance across diverse downstream tasks. Our proposed Re-Bottleneck framework offers a flexible, post-hoc solution to this limitation. By applying targeted latent-space losses, demonstrated throug...

    work page

  6. [6]

    Soundstream: An end- to-end neural audio codec,

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi, “Soundstream: An end- to-end neural audio codec,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 30, pp. 495–507, 2021

    work page 2021

  7. [7]

    High fidelity neural audio compression,

    Alexandre D ´efossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi, “High fidelity neural audio compression,”Transactions on Machine Learning Research, 2023

    2023

  8. [8]

    High-fidelity audio compression with improved rvqgan,

    Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, and Kundan Kumar, “High-fidelity audio compression with improved rvqgan,”Proc. NeurIPS, 2024

    work page 2024

Show all 36 references
  1. [9]

    Stable audio open,

    Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons, “Stable audio open,” inProc. ICASSP. IEEE, 2025, pp. 1–5

    2025

  2. [10]

    Musiclm: Generating music from text,

    Andrea Agostinelli, Timo I Denk, Zal ´an Borsos, Jesse En- gel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al., “Musiclm: Generating music from text,”arXiv preprint arXiv:2301.11325, 2023

    2023 arXiv

  3. [11]

    Vampnet: Music generation via masked acoustic token mod- eling,

    H Flores Garcia, P Seetharaman, R Kumar, and B Pardo, “Vampnet: Music generation via masked acoustic token mod- eling,” 24th Proc. ISMIR, 2023

    2023

  4. [12]

    Sound- storm: Efficient parallel audio generation,

    Zal ´an Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi, “Sound- storm: Efficient parallel audio generation,”arXiv preprint arXiv:2305.09636, 2023

    2023 arXiv

  5. [13]

    Time-domain heart sound classification using neural audio codecs,

    Ansh Mishra, Jia Qi Yip, and Eng Siong Chng, “Time-domain heart sound classification using neural audio codecs,” inProc. ICAICTA. IEEE, 2024, pp. 1–5

    2024

  6. [14]

    Speech separation using neural audio codecs with embedding loss,

    Jia Qi Yip, Chin Yuen Kwok, Bin Ma, and Eng Siong Chng, “Speech separation using neural audio codecs with embedding loss,” inProc. APSIPA ASC. IEEE, 2024, pp. 1–6

    2024

  7. [15]

    Coarse-to-fine text-to-music latent diffusion,

    Luca A Lanzend ¨orfer, Tongyu Lu, Nathana ¨el Perraudin, Dorien Herremans, and Roger Wattenhofer, “Coarse-to-fine text-to-music latent diffusion,” inProc. ICASSP. IEEE, 2025, pp. 1–5

    2025

  8. [16]

    An independence-promoting loss for music generation with language models,

    Jean-Marie Lemercier, Simon Rouard, Jade Copet, Yossi Adi, and Alexandre D ´efossez, “An independence-promoting loss for music generation with language models,” inProc. ICML, 2024

    2024

  9. [17]

    Speechtokenizer: Unified speech tokenizer for speech language models,

    Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu, “Speechtokenizer: Unified speech tokenizer for speech language models,” inProc. ICLR, 2024

    2024

  10. [18]

    Moshi: a speech-text foundation model for real- time dialogue,

    Alexandre D ´efossez, Laurent Mazar ´e, Manu Orsini, Am ´elie Royer, Patrick P ´erez, Herv ´e J ´egou, Edouard Grave, and Neil Zeghidour, “Moshi: a speech-text foundation model for real- time dialogue,” Tech. Rep., 2024

    2024

  11. [19]

    Codec-superb: An in- depth analysis of sound codec models,

    Haibin Wu, Ho-Lam Chung, Yi-Cheng Lin, Yuan-Kuei Wu, Xuanjun Chen, Yu-Chi Pai, Hsiu-Hsuan Wang, Kai-Wei Chang, Alex Liu, and Hung-Yi Lee, “Codec-superb: An in- depth analysis of sound codec models,” inFindings of ACL, 2024, pp. 10330–10348

    2024

  12. [20]

    Reconstruction vs. generation: Taming optimization dilemma in latent diffu- sion models,

    Jingfeng Yao, Bin Yang, and Xinggang Wang, “Reconstruction vs. generation: Taming optimization dilemma in latent diffu- sion models,” inProc. CVPR, 2025, pp. 15703–15712

    2025

  13. [21]

    Masked autoencoders are effective tokenizers for diffu- sion models,

    Hao Chen, Yujin Han, Fangyi Chen, Xiang Li, Yidong Wang, Jindong Wang, Ze Wang, Zicheng Liu, Difan Zou, and Bhiksha Raj, “Masked autoencoders are effective tokenizers for diffu- sion models,” inProc. ICML, 2025

    2025

  14. [22]

    Learning to upsample and upmix audio in the latent domain,

    Dimitrios Bralios, Paris Smaragdis, and Jonah Casebeer, “Learning to upsample and upmix audio in the latent domain,” arXiv preprint arXiv:2506.00681, 2025

    2025 arXiv

  15. [23]

    Beats: Audio pre-training with acoustic tokeniz- ers,

    Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, Wanxiang Che, Xiangzhan Yu, and Furu Wei, “Beats: Audio pre-training with acoustic tokeniz- ers,” inProc. ICML, 2023, pp. 5178–5193

    2023

  16. [24]

    Exploring the limits of transfer learning with a uni- fied text-to-text transformer,

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Pe- ter J Liu, “Exploring the limits of transfer learning with a uni- fied text-to-text transformer,”JMLR, 2020

    2020

  17. [25]

    Learning ordered representations with nested dropout,

    Oren Rippel, Michael Gelbart, and Ryan Adams, “Learning ordered representations with nested dropout,” inProc. ICML, 2014, pp. 1746–1754

    2014

  18. [26]

    Eq-vae: Equivariance regularized la- tent space for improved generative image modeling,

    Theodoros Kouzelis, Ioannis Kakogeorgiou, Spyros Gidaris, and Nikos Komodakis, “Eq-vae: Equivariance regularized la- tent space for improved generative image modeling,”arXiv preprint arXiv:2502.09509, 2025

    2025 arXiv

  19. [27]

    Convnext v2: Co-designing and scaling convnets with masked autoen- coders,

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie, “Convnext v2: Co-designing and scaling convnets with masked autoen- coders,” inProc. CVPR, 2023, pp. 16133–16142

    2023

  20. [28]

    The mtg-jamendo dataset for au- tomatic music tagging,

    Dmitry Bogdanov, Minz Won, Philip Tovstogan, Alastair Porter, and Xavier Serra, “The mtg-jamendo dataset for au- tomatic music tagging,” inMachine Learning for Music Dis- covery Workshop, ICML, 2019

    2019

  21. [29]

    Fma: A dataset for music analysis,

    Micha ¨el Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson, “Fma: A dataset for music analysis,” inProc. ISMIR, 2017

    2017

  22. [30]

    The song describer dataset: a corpus of audio captions for music-and- language evaluation,

    I Manco, B Weck, S Doh, M Won, Y Zhang, D Bodganov, Y Wu, K Chen, P Tovstogan, E Benetos, et al., “The song describer dataset: a corpus of audio captions for music-and- language evaluation,” NeurIPS Machine Learning for Audio Workshop, 2023

    2023

  23. [31]

    Generation or replication: Auscultating audio latent diffusion models,

    Dimitrios Bralios, Gordon Wichern, Franc ¸ois G Germain, Zexu Pan, Sameer Khurana, Chiori Hori, and Jonathan Le Roux, “Generation or replication: Auscultating audio latent diffusion models,” inProc. ICASSP. IEEE, 2024, pp. 1156– 1160

    2024

  24. [32]

    auraloss: Audio focused loss functions in pytorch,

    Christian J Steinmetz and Joshua D Reiss, “auraloss: Audio focused loss functions in pytorch,” inDigital music research network one-day workshop (DMRN+ 15), 2020

    2020

  25. [33]

    Adapting frechet audio distance for gener- ative music evaluation,

    Azalea Gui, Hannes Gamper, Sebastian Braun, and Dimitra Emmanouilidou, “Adapting frechet audio distance for gener- ative music evaluation,” inProc. ICASSP. IEEE, 2024, pp. 1331–1335

    2024

  26. [34]

    Similarity of neural network representations revisited,

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Ge- offrey Hinton, “Similarity of neural network representations revisited,” inProc. ICML, 2019, pp. 3519–3529

    2019

  27. [35]

    On kernel-target alignment,

    Nello Cristianini, John Shawe-Taylor, Andre Elisseeff, and Jaz Kandola, “On kernel-target alignment,”Proc. NeurIPS, 2001

    2001

  28. [36]

    Insights on representational similarity in neural networks with canonical correlation,

    Ari Morcos, Maithra Raghu, and Samy Bengio, “Insights on representational similarity in neural networks with canonical correlation,”Proc. NeurIPS, 2018

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.