Pith. sign in

REVIEW 3 major objections 4 minor 39 references

On a low-label soundscape task, frozen bioacoustic embeddings outperform both discrete codec tokens and general-purpose transformer embeddings under the same evaluation protocol.

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

2026-08-02 01:59 UTC pith:UVU4WLOI

load-bearing objection A solid, reproducible BirdCLEF baseline and a useful non-bird prototype trick; the headline claim about token representations is undercut by an asymmetric probe comparison. the 3 major comments →

arxiv 2607.14474 v1 pith:UVU4WLOI submitted 2026-07-16 cs.SD cs.AIcs.LG

Can Tokens Compete? Token Representations against Supervised CNN Backbones for BirdCLEF+ 2026

Anthony Miyaguchi , Murilo Gustineli , Adrian Cheung This is my paper
classification cs.SD cs.AIcs.LG
keywords BirdCLEFbioacousticssoundscape classificationaudio foundation modelsneural audio codecstransfer learninglow-resource learningprototypical networks
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.

The paper asks whether token-based audio representations—discrete codec tokens and self-supervised transformer embeddings—can compete with supervised CNN pipelines on BirdCLEF+ 2026, a multi-label soundscape task with roughly one hour of labeled audio. It first builds a supervised baseline: a frozen Perch v2 backbone read by MLP probes, blended with a trained HGNetV2-B0 sound-event-detection network and a non-bird prototypical head, reaching 0.936 on the private leaderboard within a 90-minute CPU budget. It then evaluates codec representations and general-purpose embeddings under the same frozen-probe protocol. The codec route fails badly on bioacoustic retrieval, and the semantic embeddings transfer only modestly due to domain gap; bioacoustic specialists stay far ahead. A sympathetic reading is that for low-label soundscape learning, sample-efficient probing of a domain-specialized frozen backbone beats discrete token and general-purpose representations unless those are task-adapted.

Core claim

The paper's central empirical claim is that, on a soundscape classification task with roughly one hour of labeled data, the strongest practical system is an ensemble built on a frozen bioacoustic backbone (Perch v2) probed with MLP heads and extended by a prototypical head that scores non-bird species by cosine difference to averaged positive/negative prototypes. This ensemble reaches a private leaderboard score of 0.936. The paper's comparative claim is that token representations cannot compete under the same protocol: a speech-trained neural codec (WavTokenizer) collapses Perch-embedding cosine similarity to 0.313 against a 0.90 gate, and general-purpose self-supervised transformer encoder

What carries the argument

The prototypical head: for each non-bird species, it averages the frozen Perch embeddings of labeled windows where that species is present (μ+) and absent (μ−), then scores each window by clip(cos(e_t, μ+) − cos(e_t, μ−), 0, 1), merged with the SED output via element-wise maximum. This extends a bird-only backbone to non-bird taxa and adds almost no inference cost. The comparison protocol is equally load-bearing: every encoder is evaluated by fitting a ridge head on frozen embeddings using the focal recordings, which is what determines the token-versus-specialist ranking.

Load-bearing premise

The comparison assumes that evaluating each encoder with a frozen ridge head on the focal recordings—rather than adapting token encoders to the labeled hour—captures what it means for token representations to 'compete'; the codec verdict also rests on a single speech-trained codec.

What would settle it

Re-run the comparison with the labeled soundscape hour used to adapt the token encoders (e.g., LoRA on the 66 labeled files) while keeping the frozen-probe protocol for the specialists; if any token encoder then reaches or exceeds the 0.936 private score, the conclusion that tokens cannot compete collapses.

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

If this is right

  • A frozen Perch v2 backbone probed with MLP heads scores 0.920 on the public leaderboard; adding the trained HGNetV2-B0 SED member raises the blend to 0.929, and the non-bird prototype head lifts the private score to 0.936, all within the 90-minute CPU limit.
  • The codec representation fails every predefined gate: round-trip Perch cosine is 0.313 against a 0.90 target, round-trip Precision@1 is 0.09 against 0.80, and focal retrieval Precision@10 is 0.195 against 0.80, so reconstruction-oriented tokens are not competitive for bioacoustic retrieval.
  • The offline LOSO ranking of encoders does not predict leaderboard standing—BirdNET overtakes BEATs despite lower offline AUROC—so evaluation protocol matters for transfer conclusions.
  • Most SED gains come from frontend and architecture choices (dB-scale mel normalization, higher mel resolution, backbone swap) rather than training tricks; 20+ single-variable submissions show offline AUROC anti-correlated with the public leaderboard.
  • If the paper is right, the practical route to improved soundscape models is not discrete tokenization but sequence modeling over full-band 1-D token streams, a direction the authors flag as the natural next step.

Where Pith is reading between the lines

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

  • If the pattern holds, the 'token' research direction for soundscapes should target sequence models over full-band 1-D token streams (128×1 columns) rather than codec reconstruction; the paper's Figure 4 points that way but its own SSAMBA run is inconclusive at 55k clips.
  • The one-hour labeled set shifts the winning recipe from self-supervision on focal audio to sample-efficient probing of a specialist backbone; a testable extension is whether the gap persists when the labeled set grows, where adaptive finetuning of transformer encoders may catch up.
  • The prototype head's success on non-bird taxa suggests a general recipe: extend any frozen specialist embedding to out-of-vocabulary classes by positive/negative prototype differences, but the paper only demonstrates this for Perch on the 47 non-bird species present in the labels.
  • The codec comparison used a single speech-trained codec; a bioacoustic-trained codec (or one with more tokens per second) could plausibly change the outcome, and the paper's per-species table already shows low-frequency vocalizations survive round-trip better.

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

3 major / 4 minor

Summary. The paper reports the DS@GT ARC team's entry to BirdCLEF+ 2026. The first contribution is a supervised baseline that ensembles a frozen Perch v2 backbone read out by MLP probes, a trained HGNetV2-B0 SED network, and a non-bird prototypical head, reaching a private leaderboard score of 0.936 at rank 1894 within a 90-minute CPU budget. The second and title contribution is a comparison of token-based representations against this CNN baseline: codec tokens from WavTokenizer-large are evaluated via Perch round-trip and retrieval tasks, and semantic encoders (BEATs, AST, EAT, SSAMBA, plus Perch and BirdNET as anchors) are scored with a ridge probe fit on frozen focal-recording embeddings. The paper concludes that the codec route fails on bioacoustic retrieval and that semantic embeddings transfer only modestly due to domain gap.

Significance. If the comparative claims are established, this is a useful negative result for the BirdCLEF-style low-label soundscape setting: it would show that frozen bioacoustic specialist features plus supervised probes outperform general-purpose codec and transformer token representations without task-specific fine-tuning. The baseline itself is competently engineered, with careful ablations, leaderboard verification, and a public repository; the non-bird prototype head is a particularly interesting sample-efficient idea supported by a large offline gain (macro AP 0.895 vs 0.074). The paper also ships a reproducible experimental record (Appendix A, D, E), which is a genuine strength.

major comments (3)
  1. [Table 6] The central negative claim about semantic embeddings rests on an asymmetric evaluation protocol. In Table 6, every encoder—including Perch—is scored with a ridge head fit on focal training recordings, whereas the Perch member of the baseline (§4.1) uses MLP probes fit on the 59 fully labeled soundscapes via 5-fold OOF. Appendix C shows that simply changing the probe training set lifts Perch from 0.620 (focal-trained ridge, LOSO) to 0.749 (labeled-soundscape ridge with empirical-Bayes prior), a +0.129 shift that exceeds most of the inter-encoder gaps in Table 6. Thus the conclusion 'semantic embeddings transferred only modestly due to domain gap' is not established: the comparison confounds encoder representation quality with the supervision used for the probe. Please either evaluate all encoders under the labeled-soundscape supervised protocol (e.g., MLP probes with GroupKFold OOF) or re
  2. [Table 4] The codec conclusion is based on a single neural audio codec, WavTokenizer-large, which the authors note was trained on human speech. The 'predefined success gates' in Table 4 appear to be defined for this experiment rather than drawn from an external benchmark, and no second codec (e.g., EnCodec or SoundStream) or domain-adapted codec is tested. As a result, the claim that 'the codec route failed on bioacoustic retrieval' is too broad. A single speech-trained codec failing on high-frequency bird calls is interesting but does not support a generalization about codec representations as a family. Please add at least one additional codec or substantially weaken the conclusion.
  3. [§6.1] The geometric analysis supporting the leaderboard ordering is shallow. Figure 3 uses eight hand-selected species, and the kNN accuracy is computed over recording centroids used in the same embedding plot. The text claims 'the leaderboard ordering is partly recoverable' but no confidence intervals, species sampling rationale, or comparison to random centroids is given. This does not block the baseline contribution, but it should be either removed as a supporting claim or replaced with a systematic evaluation over the 206 focal species with statistical error bars.
minor comments (4)
  1. [Conclusion] Typo: 'brough' should be 'brought'.
  2. [Table 6] The caption says 'AUROC is the leaderboard-aligned score from the 9-fold LOSO ridge probe' but Table 2 reports 'macro AUROC' for the baseline. It would help to explicitly state that the Table 6 Perch row is the ridge-probe configuration (0.866 public), not the MLP-probe baseline (0.920), to avoid confusion with §4.
  3. [Figure 2] The caption states 'Public LB scores are shown in green,' but the figure is grayscale in the PDF; if color is used, ensure it is distinguishable in print or add numerical labels.
  4. [Appendix D, Table 14] The reported 'Full BEATs (teacher)' accuracy of 84.0% seems inconsistent with Table 13, where BEATs linear probe obtains 95.8% on the same ESC-50 task; clarify which setup produces the teacher number and why it is lower.

Circularity Check

1 steps flagged

Peripheral in-sample prototype evaluation; central benchmark claims are externally grounded

specific steps
  1. fitted input called prediction [Section 4.1 (Non-Bird Prototypical Head) and Section 4.2 (Leaderboard Results)]
    "Of the 72 non-bird taxa, 47 appear in the labeled soundscapes, allowing us to construct positive (μ+s) and negative (μ−s) prototypes by averaging the Perch embeddings of all windows where the species is present, and all windows where it is absent, respectively. ... Offline evaluation on the 66 labeled soundscapes shows the prototype substantially outperforms the SED model alone for non-bird taxa (macro AP 0.895 vs. 0.074)."

    The prototypes are computed as the mean Perch embedding over every labeled soundscape window for each species, and the reported 'offline evaluation' is on the same 66 labeled soundscapes. The proto score is thus a nearest-centroid rule whose centroids include the very windows being scored, so the 0.895 macro AP is a training-set fit rather than a held-out prediction. It is not independent evidence that the Perch embedding space 'carries useful information for non-bird species.' However, the final leaderboard gain from this head (+0.002 public, +0.011 private) is external, so this circularity is peripheral rather than load-bearing.

full rationale

The paper's main empirical claims are anchored to external benchmarks: the private LB score of 0.936 (Table 2) and ESC-50 screening (Table 13). The token-vs-CNN comparison in Table 6 uses one protocol for all six encoders, including the Perch and BirdNET anchors: 'we fit a ridge head on each encoder's frozen embeddings using the focal training recordings.' Table 2 confirms that the Perch ridge line (0.866) is the same head replicated in Table 6, so the general-encoder deficit is not an artifact of giving the baseline labeled-soundscape MLP probes while denying them to token encoders. The codec verdict rests on explicit external thresholds and Perch as an independent round-trip evaluator, not on a fitted input. The one genuinely circular passage is the offline validation of the non-bird prototypical head, as detailed in the step above; it is a training-set score and is not load-bearing because the head is also validated externally on the leaderboard. Self-citations to the team's prior BirdCLEF notes are narrative, and no uniqueness theorem or ansatz is imported from them. The paper itself notes that offline validation AUROC anti-correlated with the public leaderboard, a calibration limitation, but that is not a circularity. Overall: no significant central circularity; one peripheral in-sample evaluation.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The central score depends on standard ML training of frozen and trained networks; the comparative conclusions rest on evaluation-protocol choices: using Perch as the evaluator for codec round-trips, using ridge probes for encoders, and treating WavTokenizer as representative of codecs. No new physical or architectural entities are postulated; all components are adapted from prior work. Learned network weights themselves are omitted because they are the model rather than introduced constants.

free parameters (8)
  • Ensemble weight w_Perch = 0.5
    Equal-weight blend is chosen by hand; the ensemble scores 0.929 vs 0.920/0.917 for members, and the exact weight affects the final private score.
  • Ridge regularization strength λ = not reported
    Ridge heads are fit per encoder on focal training data for Table 6; the regularization value is not given and directly affects the frozen-embedding rankings.
  • MLP probe hyperparameters = not reported
    Perch probe head trained on 59 labeled soundscapes; depth, width, epochs, and learning rate are not specified; standalone probe scores 0.920.
  • Site/hour prior parameters = not reported
    Post-processing adds +0.007 public LB; the smoothing or prior construction is not specified.
  • Prototype positive/negative embeddings = derived from labeled windows
    The non-bird head is built by averaging Perch embeddings across labeled soundscape windows; this is a data-fitted component central to the 0.936 score.
  • WavTokenizer success gates = cos>0.90; P@1>0.80; P@10>0.80
    Table 4 calls these 'predefined' without a source; they are hand-set thresholds. The failure margins are large enough that the qualitative conclusion is robust, but the gates are unverified.
  • k in kNN class-separability = 10
    Hand-chosen for the kNN ranking in §6.1; illustrative rather than load-bearing.
  • LoRA rank = 8
    Used for EAT and SSAMBA in ESC-50 appendix experiments; not used in the main frozen-embedding LOSO ranking.
axioms (6)
  • domain assumption The Kaggle leaderboard macro-AUROC is a valid and reliable measure of target-task performance.
    All central scores are self-reported from private/public LB; no submission artifacts or error bars are provided.
  • domain assumption Frozen Perch v2 embeddings carry transferable structure for non-bird taxa.
    The prototype head relies on Perch (bird-trained) to cluster frogs, insects, and mammals; offline macro AP 0.895 supports this on this dataset, but it is not guaranteed generally.
  • ad hoc to paper WavTokenizer-large is representative of neural audio codecs for this comparison.
    Only one speech-trained codec is tested (Table 4); the conclusion 'codec route failed' generalizes beyond the evidence.
  • domain assumption A frozen ridge probe is a fair way to rank encoders on the target task.
    Token encoders are not adapted to the labeled soundscapes in Table 6; Appendix D shows LoRA can change ESC-50 accuracy dramatically, so rankings might change with adaptation.
  • ad hoc to paper Success gates in Table 4 are pre-specified.
    The paper calls them 'predefined' but gives no reference or pre-registration; if chosen post hoc they weaken the codec-failure test.
  • standard math GroupKFold/LOSO splits and PaCMAP provide unbiased validation and visualization.
    Standard statistical procedures; no unusual assumptions beyond typical cross-validation and dimensionality-reduction behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 15473 in / 15900 out tokens · 148907 ms · 2026-08-02T01:59:07.149121+00:00 · methodology

0 comments
read the original abstract

This paper details the DS@GT ARC team's approach to BirdCLEF+ 2026, multi-label detection of animal vocalizations in soundscapes from the Pantanal wetlands. The 2026 edition adds about an hour of labeled soundscapes, shifting the task toward supervised pipelines fit to the labeled set. First, we build a competitive supervised baseline that ensembles a frozen Perch v2 backbone, a trained HGNetV2-B0 sound-event-detection network, and a non-bird prototypical head, reaching a private leaderboard score of 0.936 at rank 1894 within a 90-minute CPU budget. Second, we ask whether token-based representations can compete, contrasting codec representations from neural audio codecs against semantic representations from foundational embeddings. We compare two bioacoustic specialist models against four token-based encoders trained on AudioSet. The repository for this work can be found at https://github.com/dsgt-arc/birdclef-2026.

Figures

Figures reproduced from arXiv: 2607.14474 by Adrian Cheung, Anthony Miyaguchi, Murilo Gustineli.

Figure 1
Figure 1. Figure 1: Representative spectrograms of three vocalizing taxa in BirdCLEF 2026. The taxa overlap in frequency. Frogs produce pulsed call trains, birds frequency-modulated whistles, and insects a sustained broadband droning sound [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The supervised baseline is an equal-weight, two-member ensemble extended with a non-bird prototypical head. Each 60 s soundscape is split into 12 non-overlapping 5 s windows that feed both members: a frozen Perch v2 backbone read through logit-space MLP probes, and a trained HGNetV2-B0 SED network on log-mel spectrograms. Their 234-species probability vectors are blended equally, post-processed with site/h… view at source ↗
Figure 3
Figure 3. Figure 3: Embedding-space species separability across the six deployed encoders. Two-dimensional PaCMAP projections of frozen embeddings for eight BirdCLEF 2026 species (one point per focal recording, mean-pooled). To analyze a feature encoder, we visualize how classes cluster in space. Our results are consistent with [6], where birdsong embeddings out-compete AudioMAE and show better separability in a t-SNE scatter… view at source ↗
Figure 4
Figure 4. Figure 4: Patch geometry on ViT-styled tokenization of spectrograms. (a) The default ViT grid ties tokens to a fixed 2-D layout where audio is locked into dimensions of the training data. (b) Full-band 128 × 1 columns collapse the entire spectrum into each token (128 mel bins × 10 ms, one per frame) yielding a 1-D stream. (c) Because the stream is one-dimensional, a sequence model can contextualize an entire soundsc… view at source ↗
Figure 5
Figure 5. Figure 5: Spectral band occupancy by taxon, over 80 random recordings each. Left: mean energy-normalized power spectrum. Right: per-recording peak frequency, with the median marked. All three taxa concentrate energy in 1–4 kHz (peak-frequency medians: frogs 1.8, birds 2.9, insects 1.8 kHz), so frequency band alone does not separate them. C. Frozen Perch Transfer Learning This appendix documents exploratory probes on… 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

39 extracted references · 12 linked inside Pith

  1. [1]

    Overview of

    Picek, Lukas and Adam, Luk. Overview of. International Conference of the Cross-Language Evaluation Forum for European Languages (CLEF) , year=

  2. [2]

    Overview of

    Kahl, Stefan and Denton, Tom and Sugai, Larissa and Piatti, Litiana and Arruda Moreno, Wener Hugo and Barbosa, Mariana Motti and Eibl, Maximilian and Fieker, Carolline Zatta and Garcia, Carolina Martins and Hokama Sousa, Daiene Louveira and J. Overview of. Working Notes of CLEF 2026 -- Conference and Labs of the Evaluation Forum , year=

    2026

  3. [3]

    Advances in neural information processing systems , volume=

    wav2vec 2.0: A framework for self-supervised learning of speech representations , author=. Advances in neural information processing systems , volume=

  4. [4]

    Overview of

    Ca. Overview of. Working

  5. [5]

    Proceedings of the 40th

    Chen, Sanyuan and Wu, Yu and Wang, Chengyi and Liu, Shujie and Tompkins, Daniel and Chen, Zhuo and Che, Wanxiang and Yu, Xiangzhan and Wei, Furu , year = 2023, series =. Proceedings of the 40th. 2212.09058 , pages =

    Pith/arXiv arXiv 2023

  6. [6]

    Proceedings of the

    Chen, Wenxi and Liang, Yuzhe and Ma, Ziyang and Zheng, Zhisheng and Chen, Xie , year = 2024, eprint =. Proceedings of the

    2024

  7. [7]

    D. High. 2210.13438 , archiveprefix =

    Pith/arXiv arXiv

  8. [8]

    2021 , eprint=

    Improving Bird Classification with Unsupervised Sound Separation , author=. 2021 , eprint=

    2021

  9. [9]

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , year = 2021, eprint =. An. International

    2021

  10. [10]

    and Ellis, Daniel P

    Gemmeke, Jort F. and Ellis, Daniel P. W. and Freedman, Dylan and Jansen, Aren and Lawrence, Wade and Moore, R. Channing and Plakal, Manoj and Ritter, Marvin , year = 2017, month = mar, pages =. Audio. 2017

    2017

  11. [11]

    Scientific Reports , volume=

    Global birdsong embeddings enable superior transfer learning for bioacoustic classification , author=. Scientific Reports , volume=. 2023 , publisher=

    2023

  12. [12]

    Mind the

    Liang, Jinhua and Nolasco, Ines and Ghani, Burooj and Phan, Huy and Benetos, Emmanouil and Stowell, Dan , year = 2024, month = mar, eprint =. Mind the

    2024

  13. [13]

    doi:10.21437/Interspeech.2021-698 , archiveprefix =

    Interspeech 2021 , author =. doi:10.21437/Interspeech.2021-698 , archiveprefix =. 2104.01778 , pages =

    Pith/arXiv arXiv 2021

  14. [14]

    Gu, Albert and Dao, Tri , year = 2023, month = dec, eprint =. Mamba:

    2023

  15. [15]

    Hagiwara, Masato , year = 2022, month = oct, eprint =

    2022

  16. [16]

    Perch 2.0:

    van Merri. Perch 2.0:. 2508.04665 , archiveprefix =

    arXiv

  17. [17]

    and Kennard, Robert W

    Hoerl, Arthur E. and Kennard, Robert W. , year = 1970, month = feb, journal =. Ridge

    1970

  18. [18]

    and Shen, Yelong and Wallis, Phillip and

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and. International. 2106.09685 , archiveprefix =

    Pith/arXiv arXiv

  19. [19]

    Ji, Shengpeng and Jiang, Ziyue and Wang, Wen and Chen, Yifu and Fang, Minghui and Zuo, Jialong and Yang, Qian and Cheng, Xize and Wang, Zehan and Li, Ruiqi and Zhang, Ziang and Yang, Xiaoda and Huang, Rongjie and Jiang, Yidi and Chen, Qian and Zheng, Siqi and Zhao, Zhou , year = 2024, month = aug, eprint =

    2024

  20. [20]

    and Eibl, Maximilian and Klinck, Holger , year = 2021, month = mar, journal =

    Kahl, Stefan and Wood, Connor M. and Eibl, Maximilian and Klinck, Holger , year = 2021, month = mar, journal =

    2021

  21. [21]

    Overview of

    Kahl, Stefan and Navine, Amanda and Denton, Tom and Klinck, Holger and Hart, Patrick and Glotin, Herv. Overview of. Proceedings of the

  22. [22]

    Overview of

    Kahl, Stefan and Denton, Tom and Klinck, Holger and Reers, Hendrik and Cherutich, Francis and Glotin, Herv. Overview of. Working

  23. [23]

    Overview of

    Kahl, Stefan and Denton, Tom and Klinck, Holger and Ramesh, Vijay and Joshi, Viral and Srivathsa, Meghana and Anand, Akshay and Arvind, Chiti and Cp, Harikrishnan and Sawant, Suyash and Vv, Robin and Glotin, Herv. Overview of. Working

  24. [24]

    Liu, Zhuang and Mao, Hanzi and Wu, Chao-Yuan and Feichtenhofer, Christoph and Darrell, Trevor and Xie, Saining , year = 2022, eprint =. A. Proceedings of the. doi:10.1109/CVPR52688.2022.01167 , archiveprefix =

    arXiv 2022

  25. [25]

    Efficient

    Mikolov, Tomas and Chen, Kai and Corrado, Greg and Dean, Jeffrey , year = 2013, eprint =. Efficient. International

    2013

  26. [26]

    Miyaguchi, Anthony and Yu, Jiangyue and Cheungvivatpant, Bryan and Dudley, Dakota and Swain, Aniketh , editor =. Motif. Proceedings of the. 2206.04805 , pages =

    Pith/arXiv arXiv

  27. [27]

    Transfer

    Miyaguchi, Anthony and Zhong, Nathan and Gustineli, Murilo and Hayduk, Chris , editor =. Transfer. Working. 2306.16760 , pages =

    Pith/arXiv arXiv

  28. [28]

    Transfer

    Miyaguchi, Anthony and Cheung, Adrian and Gustineli, Murilo and Kim, Ashley , editor =. Transfer. Working. 2407.06291 , pages =

    Pith/arXiv arXiv

  29. [29]

    Distilling

    Miyaguchi, Anthony and Gustineli, Murilo and Cheung, Adrian , editor =. Distilling. Working. 2507.08236 , pages =

    Pith/arXiv arXiv

  30. [30]

    , year = 2015, month = oct, pages =

    Piczak, Karol J. , year = 2015, month = oct, pages =. Proceedings of the 23rd

    2015

  31. [31]

    Hopfield

    Ramsauer, Hubert and Sch. Hopfield. International. 2008.02217 , archiveprefix =

    Pith/arXiv arXiv 2008

  32. [32]

    Rauch, Lukas and Heinrich, Ren. Can. 2504.12880 , archiveprefix =

    Pith/arXiv arXiv

  33. [33]

    Rauch, Lukas and Schwinger, Raphael and Wirth, Moritz and Heinrich, Ren. The. 2403.10380 , publisher =

    Pith/arXiv arXiv

  34. [34]

    Foundation

    Schwinger, Raphael and Zadeh, Paria Vali and Rauch, Lukas and Kurz, Mats and Hauschild, Tom and Lapp, Sam and Tomforde, Sven , year = 2025, month = aug, eprint =. Foundation

    2025

  35. [35]

    Shams, Siavash and Dindar, Sukru Samet and Jiang, Xilin and Mesgarani, Nima , year = 2024, month = dec, pages =

    2024

  36. [36]

    Advances in neural information processing systems , volume=

    Prototypical networks for few-shot learning , author=. Advances in neural information processing systems , volume=

  37. [37]

    , year = 2019, series =

    Tan, Mingxing and Le, Quoc V. , year = 2019, series =. Proceedings of the 36th. 1905.11946 , pages =

    Pith/arXiv arXiv 2019

  38. [38]

    Tang, Chengyu and Baskiyar, Sanjeev , year = 2025, month = dec, eprint =. State

    2025

  39. [39]

    Zeghidour, Neil and Luebs, Alejandro and Omran, Ahmed and Skoglund, Jan and Tagliasacchi, Marco , year = 2021, month = jul, eprint =

    2021