Pith. sign in

REVIEW 4 major objections 6 minor 28 references

The paper puts forward a unified framework that recasts automatic Roman-numeral harmony analysis as a spectrum from blind full-score prediction to partial completion to analyst-constrained revision, and supports all three through one hybrid

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 04:43 UTC pith:K2MPGW2K

load-bearing objection Real modeling contribution with an honest but unvalidated interactive claim: masked completion works, but the analyst-in-the-loop story needs direct testing before the headline sells it. the 4 major comments →

arxiv 2607.13587 v1 pith:K2MPGW2K submitted 2026-07-15 cs.SD cs.AI

From Prediction to Collaboration: Interactive Symbolic Music Analysis

Emmanouil Karystinaios , Johannes Hentschel , Markus Neuwirth , Gerhard Widmer This is my paper
classification cs.SD cs.AI
keywords Roman numeral analysissymbolic music analysisinteractive analysismasked completiongraph neural networkspretrained music encodersmulti-task learningdistillation
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 argues that automatic Roman-numeral harmony analysis should be treated as a continuum rather than a one-shot prediction task. It introduces a single framework that handles blind full-score inference, completion of missing labels from partial context, and constrained revision after analyst edits, by combining a pretrained symbolic encoder with a graph neural network. The key evidence is a masked-completion experiment on a heterogeneous benchmark: accuracy at predicting held-out labels rises monotonically from about 58% with only 5% of labels known to about 83% with 95% known, overtaking blind inference before half the labels are supplied. The paper also introduces a distilled edit-conditioned model that enforces hard constraints on known labels and a prototype interface for multi-level candidate inspection. Notably, the authors stop short of claiming demonstrated end-user benefit; they frame the contribution as the modeling and system components needed for interactive workflows.

Core claim

The central claim is that one modeling stack can serve all three modes of use: blind full-score prediction, partial-label completion, and constrained revision. The full RNHybrid model couples pretrained contextual note representations with a multi-task graph network over note, onset, beat, and measure structures; the masked edit-conditioned variant is distilled from that model and injects known labels as hard constraints during inference. On the Dilemmadata benchmark, masked completion accuracy improves monotonically as the known-label ratio grows, reaching 0.831 on AugNet and 0.801 on DLC at 95% context, with low variance across mask samples. The authors also show that inference-time decodi

What carries the argument

The mechanism that carries the argument is the edit-conditioned masked checkpoint, distilled from a frozen teacher. Known labels are embedded and injected as a note-level conditioning bias before graph message passing; unknown positions receive a masked embedding; and at inference time, known labels are enforced as hard constraints. Training mixes masked and unmasked batches with a hybrid policy of random-node and onset-span masking, supported by three preservation losses (KL distillation, feature anchoring, and L2-SP drift penalty). The full model pairs this with a pretrained symbolic encoder (MusicBERT-style) feeding a multi-task graph network, plus post-hoc aggregation voters and beam dec

Load-bearing premise

The masked checkpoint was distilled from the paper's own full-inference model and tested only with randomly sampled masks, so the claimed support for analyst-in-the-loop revision rests on the untested assumption that the model behaves the same under realistic, ordered, localized human edits.

What would settle it

Run the masked-completion model on masks that mimic analyst behavior—for example, corrections concentrated in one contiguous phrase, or labels that overturn the model's top prediction—and compare accuracy on the remaining positions against blind inference. If accuracy does not stay at or above the random-mask curve, the interactive benefit claimed in the paper is not established.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single system can serve blind analysis, partial annotation completion, and post-edit re-prediction, removing the need for task-specific models.
  • Because masked accuracy exceeds blind inference before 50% of labels are known, sparse human annotations could reliably steer predictions on difficult passages.
  • Hard constraints guarantee that analyst-confirmed labels are never overwritten, which is a minimal requirement for trust in iterative tools.
  • The multi-level candidate interface (note, onset, beat, measure) turns per-note probabilities into the objects analysts actually read, enabling auditable disagreement.
  • The framework sets up the next validation step: measuring whether the components improve real analyst workflows in speed and correction behavior.

Where Pith is reading between the lines

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

  • If the masked model generalizes from random masks to edit patterns (contiguous regions, error-dense spots), the 50% crossover could support an active-learning loop that asks analysts to annotate only the highest-uncertainty passages; this is a natural extension the paper does not test.
  • The post-hoc decoders being neutral or harmful under blind inference suggests the bottleneck for interactive use is not the decoder but the training and evaluation regime; a direct comparison of those decoders under edit-conditioned inference would separate architecture gains from regime gains.
  • The paper's monotone completion curve is measured on randomly sampled masks; a sterner test would be edits that deliberately contradict the model's prior, since a useful tool must recover from an analyst correcting a label the model was confident about.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents RNHybrid, a hybrid MusicBERT+graph model for symbolic Roman-numeral (RN) analysis, together with a distilled masked edit-conditioned checkpoint for partial-label completion and a Verovio-based interface for multi-level candidate inspection and editing. The framework is evaluated on the authors' Dilemmadata benchmark (AugNet and DLC subsets), with experiments covering blind full-inference accuracy, masked completion accuracy at varying known-label ratios, multitask conflict-handling optimizers, and post-hoc decoding modules. The central claim is that a single framework can support blind analysis, constrained completion, and iterative revision; however, the interactive/revision component is not directly evaluated, and several evaluation gaps weaken the empirical case.

Significance. If the interactive capabilities were properly validated, the paper would constitute a useful step toward human-in-the-loop symbolic music analysis, a direction that is relatively underexplored in the RN literature. The hybrid architecture yields competitive blind inference on two corpora, and the monotonic masked-completion trend in Table 2 is a novel and potentially useful empirical finding. The paper is transparent about its limitations, explicitly disclaiming user-centered validation, and it releases code. These are strengths. Nevertheless, the evidence for the core interactive/revision claim is indirect, the evaluation is self-referential (self-cited benchmark, self-distilled teacher), and key tables lack error bars or significance tests. The current results support a claim of 'partial-label completion under random masks', not the stronger 'interactive revision' contribution advertised in the title and abstract.

major comments (4)
  1. [Secs. 3.3, 6, Table 4] The central interactive-revision claim is not directly tested. The masked edit-conditioned checkpoint (Sec. 3.3) is evaluated only with randomly sampled masks (Sec. 4.2), and the post-hoc decoders (voter, iterative refinement, beam search) are evaluated only under blind full inference in Table 4, where they are neutral or harmful. No experiment combines the edit-conditioned model with these decoders or uses realistic edit patterns (ordered, localized corrections). The paper acknowledges this in Sec. 4.2 and disclaims user-centered validation in Secs. 6-7, but the system-level claim of supporting constrained revision (Abstract, Sec. 1, Sec. 6) still relies on this untested transfer. Please either add an edit-conditioned evaluation (e.g., masks derived from analyst-like corrections, decoders applied in that regime) or substantially soften the claims to 'partial-label completion' only.
  2. [Table 2] Table 2 lacks a no-conditioning baseline that controls for which positions are masked at each known-label ratio. At 5% known labels, the AugNet masked accuracy (0.577) is statistically identical to the blind full-inference RN accuracy (0.576, Table 1), so the monotonic improvement could be partly due to the changing composition of the masked set (e.g., smaller masked sets containing easier positions) rather than to effective use of context. A baseline that predicts the same masked positions using the blind model, or a context-only baseline (e.g., nearest-neighbor labels), is needed to attribute the improvement to conditioning.
  3. [Sec. 4.1, ref [5]] All experiments use Dilemmadata, a benchmark introduced by two of the authors in a separate paper (ref [5]) that is not described here. The test subsets are said to be 'curated to avoid data-leakage' (Sec. 4.1), but the split, leakage checks, and annotation harmonization are not specified. Because the central empirical claims depend on this benchmark, the lack of detail and the self-referential provenance (masked student distilled from the authors' own teacher) limit external validation. Please provide the benchmark details or make the dataset and preprocessing code publicly available, and discuss any risk of circularity.
  4. [Tables 1, 3, 4] Tables 1, 3, and 4 report point estimates without error bars or significance tests. For example, in Table 1 (AugNet RN), RNHybrid scores .576 vs. RNBert .574, a 0.2-point difference; in Table 4, iterative refinement improves RN accuracy by +0.0092 while increasing loss by +0.5310. These differences are likely within run-to-run noise. Without multiple seeds or statistical tests, the claims that RNHybrid is 'competitive' and that the decoders are 'neutral or harmful' are not grounded. Please add error bars or significance tests, or at least state the number of seeds.
minor comments (6)
  1. [Abstract, Secs. 4.1, 6] Typographical errors: 'Percepetion' (affiliations), 'Brukner' (for Bruckner), 'Adressing' (Sec. 6), 'the the DLC testset' (Table 3 caption).
  2. [Abstract] The claim that Dilemmadata is 'the largest and most heterogeneous benchmark of its kind' is not substantiated in the paper; since the benchmark is only cited, please provide details or temper the claim.
  3. [Sec. 5] The text states masked completion 'passes blind full inference before the 50% mark', but Table 2 does not include the blind full-inference value for direct comparison. Include the blind value in the same table or specify exactly which number is being compared.
  4. [Sec. 4.2] The statement 'we additionally inspect note-level RN accuracy and observe no substantial difference on our validation runs' is vague; report the actual numbers or remove the sentence.
  5. [Figure 2] The interface is described only briefly; a more detailed walkthrough (e.g., how the top-3 candidates are computed, how edits propagate through the masked model) would improve reproducibility of the system design.
  6. [References] Ref [16] is a HuggingFace model checkpoint, not a peer-reviewed paper; cite it as software/online resource in the appropriate format.

Circularity Check

0 steps flagged

No significant circularity: predictions are evaluated against external ground-truth labels, and self-citations supply components/data rather than forcing the derivational conclusions.

full rationale

The paper's central claims are empirical and are not obtained by definitional reduction or by fitting inputs renamed as predictions. The masked edit-conditioned checkpoint is distilled from the paper's own full-inference model, but Table 2 accuracy is computed on masked positions against the external reference annotations of Dilemmadata/AugmentedNet/DC, not against the teacher or the training targets; distillation is a training mechanism, not the evaluation metric. The observed monotonic improvement with more known labels is a measured model behavior, not an identity. The benchmark Dilemmadata is cited from the authors' own group, and several architectural components (AnalysisGNN, ChordGNN, GraphMuse, the MusicBERT checkpoint) are also self-cited, but these citations supply a dataset, code, or pretrained weights; they do not encode the paper's result, and the core comparison includes published external baselines. The paper itself flags the key limitation: Section 4.2 states 'Known labels are selected according to the masking procedure described above rather than from recorded analyst corrections,' and Section 7 disclaims 'the end-user benefits of a complete analyst-facing workflow.' That is a missing validation of transfer to real analyst edits, not a circular reduction. No uniqueness theorem, ansatz, or fitted parameter is imported via self-citation to force the conclusions. Thus no circular step can be exhibited, and the derivation chain is self-contained in the relevant sense.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 1 invented entities

The central claims rest on a shared task schema inherited from prior work, a self-created benchmark, and a masked student distilled from the authors' own full model. These are reasonable for an ML systems paper, but they mean that the strongest evidence for interactive utility is not yet independent of the authors' own choices.

free parameters (1)
  • Learned model weights (RNHybrid and masked student) = not enumerated
    All network weights are optimized on training splits of Dilemmadata. No additional hand-set constants are introduced for the main claims beyond standard hyperparameters such as masking ratios and LoRA rank, which are not reported with sensitivity analysis.
axioms (4)
  • domain assumption Roman-numeral labels decompose into local key, degree, quality, and inversion
    The task schema follows prior RN work (refs [2,3,6,15]) and conditions what the model predicts. If this decomposition fails for some repertoire (e.g., modal or extended tonal), the framework inherits the error.
  • domain assumption The harmonized Dilemmadata labels are accurate ground truth and its splits avoid leakage
    Section 4.1 treats the new benchmark as authoritative. Since benchmark [5] is from two of the present authors, this is not independently audited within this paper.
  • ad hoc to paper A distilled masked student with frozen MusicBERT embeddings is sufficient for edit-conditioned completion
    The paper adopts a distillation/masking recipe rather than a from-scratch conditional model. Table 2 shows absolute performance but no comparison to alternate formulations, so the sufficiency of this design is asserted.
  • ad hoc to paper Post-hoc decoding modules are useful in edit-conditioned decoding despite being neutral or harmful under blind inference
    Table 4 shows these modules do not help blind inference. Section 6 asserts their proper regime is edit-conditioned, but no edit-conditioned experiment is provided, making this an unverified premise for the interactive claim.
invented entities (1)
  • Dilemmadata benchmark no independent evidence
    purpose: Evaluation benchmark combining AugmentedNet and Distant Listening corpora under a harmonized schema, used for all main experiments.
    The benchmark is introduced in ref [5] by two of the present authors. The underlying corpora are external, but the benchmark assembly, harmonization, and split selection are self-produced and not released or audited within this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 8844 in / 10502 out tokens · 107553 ms · 2026-08-02T04:43:05.015729+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of From Prediction to Collaboration: Interactive Symbolic Music Analysis." pith.science (2026). https://pith.science/paper/K2MPGW2K

@misc{pith2026260713587,
  author       = {Pith},
  title        = {Pith review of: From Prediction to Collaboration: Interactive Symbolic Music Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2MPGW2K}},
  note         = {Machine review of arXiv:2607.13587}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automatic symbolic music analysis has made substantial progress, yet existing systems are typically designed for a single mode of use, such as full-score prediction, and therefore do not match the broader range of operations that arise in analysis workflows, including partial completion, local correction, and iterative refinement. As a result, there remains a gap between strong benchmark models and systems that can support interactive analytical use. We present a unified framework for symbolic Roman-numeral (RN) analysis that narrows this gap by combining strong predictive performance with direct support for constrained completion and revision. The method is designed to provide a practical trade-off between accuracy and interactive responsiveness by computing expensive pretrained representations once and reusing them during iterative refinement, making powerful pretrained models more amenable to interactive settings. It supports complete score analysis, targeted revision of existing labels, and inference of missing annotations from partial context through a shared modeling framework. Experiments on Dilemmadata, the largest and most heterogeneous benchmark of its kind, show that the proposed approach is a strong RN-analysis baseline while also supporting masked completion from partial labels. Together with a prototype interface for multi-level candidate inspection and editing, these results position automatic RN analysis not only as a prediction problem, but also as a foundation for future interactive tools for music analysis.

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

28 extracted references · 2 canonical work pages

  1. [1]

    From Prediction to Collaboration: Interac- tive Symbolic Music Analysis

    INTRODUCTION Roman numeral (RN) analysis describes chords and har- monic functions relative to a local key. It is widely used for Western tonal music and can also be applied to some modal and extended-tonal repertoire. Automatic RN analy- sis remains challenging because harmonic labels are highly context-dependent and must stay coherent across multiple mu...

    Pith/arXiv arXiv 2026

  2. [2]

    Early systems used recurrent and multi-task formulations [8, 9], while later attention-based models improved the han- dling of longer-range harmonic context [10]

    RELATED WORK Prior work on automatic Roman numeral analysis has largely focused on improving blind full-score prediction. Early systems used recurrent and multi-task formulations [8, 9], while later attention-based models improved the han- dling of longer-range harmonic context [10]. Augmented- Net further strengthened this line of work by combining multi...

  3. [3]

    f: Ger6/V/i

    METHODOLOGY Our goal is to support three complementary analysis modes within one framework: (i) complete-score inference, (ii) local revision of existing labels, and (iii) completion of par- tially annotated passages. Our system combines an RNBert- inspired sequence component based on pretrained repre- sentations with an AnalysisGNN-style multi-task graph...

    2026

  4. [4]

    Post-hoc voter

    EXPERIMENTS We evaluate four regimes. First, we measure full inference performance compared to previous models and different model settings. Second, we evaluate masked completion with our edit-conditioned model, where part of the analysis is known and the model predicts the missing labels with explicit known-label preservation. Third, we evaluate dif- fer...

    2026

  5. [5]

    RESULTS The strongest current result is not a single model, but the combination of a robust full-analysis checkpoint, a dedicated masked completion checkpoint, and a flexible post-hoc inference stack. The full checkpoint provides the best blind-analysis baseline, while the masked model en- ables context-aware completion with perfect consistency on known R...

    2026

  6. [6]

    Analysts, however, do not reason about notes in isolation; they work with harmonic labels placed at change-points and scoped to beats, measures, or longer segments

    DISCUSSION A central obstacle to making neural RN analysis useful in practice is structural rather than modelling-based: symbolic scores are presented to sequence, graph, and slice-based networks alike as sets of notes, and analysis tasks are trained and evaluated as one label per task per note. Analysts, however, do not reason about notes in isolation; t...

  7. [7]

    CONCLUSION AND FUTURE WORK We presented a framework that recasts symbolic RN anal- ysis as a prediction-to-analysis continuum: a hybrid Mu- sicBERT + graph backbone that makes blind full-score infer- ence competitive with strong published baselines on Dilem- madata, paired with an edit-conditioned masked checkpoint and a Verovio-based interface that turns...

    2026

  8. [8]

    Towards a Unified Model of Musical Form: Bridging Music The- ory, Digital Corpus Research, and Computation

    ACKNOWLEDGEMENTS This research has been supported by the Swiss National Science Foundation (SNSF) through the project "Towards a Unified Model of Musical Form: Bridging Music The- ory, Digital Corpus Research, and Computation" (grant no. 10000183; 2024-2028) and by the European Research Council (ERC) under the EU’s Horizon 2020 research & innovation progr...

    2024

  9. [9]

    Musicbert: Symbolic music understanding with large- scale pre-training,

    M. Zeng, X. Tan, R. Wang, Z. Ju, T. Qin, and T.-Y . Liu, “Musicbert: Symbolic music understanding with large- scale pre-training,” inFindings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 2021, pp. 791–800

    2021

  10. [10]

    Rnbert: Fine-tuning a masked language model for roman numeral analysis,

    M. Sailor, “Rnbert: Fine-tuning a masked language model for roman numeral analysis,” inProceedings of the International Society for Music Information Re- trieval Conference (ISMIR), San Francisco, United States, 2024, pp. 814–821

    2024

  11. [11]

    Roman numeral anal- ysis with graph neural networks: Onset-wise predic- tions from note-wise features,

    E. Karystinaios and G. Widmer, “Roman numeral anal- ysis with graph neural networks: Onset-wise predic- tions from note-wise features,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), Milan, Italy, 2023, pp. 597–604

    2023

  12. [12]

    Analysisgnn: Unified music analysis with graph neural networks,

    E. Karystinaios, J. Hentschel, M. Neuwirth, and G. Wid- mer, “Analysisgnn: Unified music analysis with graph neural networks,” inProceedings of the International Symposium on Computer Music Multidisciplinary Re- search, 2025

    2025

  13. [13]

    Dilemmadata: A symbolic dataset for music research,

    J. Hentschel and E. Karystinaios, “Dilemmadata: A symbolic dataset for music research,” inProceedings of the Music Encoding Conference (MEC), 2026

    2026

  14. [14]

    AugmentedNet: A Roman Numeral Analysis Network with Synthetic Training Examples and Additional Tonal Tasks,

    N. Nápoles López, M. Gotham, and I. Fujinaga, “AugmentedNet: A Roman Numeral Analysis Network with Synthetic Training Examples and Additional Tonal Tasks,” inProceedings of the 22nd International Society for Music Information Retrieval Conference, ISMIR. Online: ISMIR, Nov. 2021, pp. 404–411. [Online]. Available: https://zenodo.org/record/5624533

    arXiv 2021

  15. [15]

    A corpus and a modular infrastructure for the empirical study of (an)notated music,

    J. Hentschel, Y . Rammos, M. Neuwirth, and M. Rohrmeier, “A corpus and a modular infrastructure for the empirical study of (an)notated music,”Scientific Data, vol. 12, no. 1, p. 685, 2025. [Online]. Available: https://doi.org/10.1038/s41597-025-04976-z

    doi:10.1038/s41597-025-04976-z 2025

  16. [16]

    Functional harmony recognition of symbolic music data with multi-task recurrent neural networks,

    T.-C. Chen and L. Su, “Functional harmony recognition of symbolic music data with multi-task recurrent neural networks,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), Paris, France, 2018, pp. 90–97

    2018

  17. [17]

    Not all roads lead to rome: Pitch representation and model archi- tecture for automatic harmonic analysis,

    G. Micchi, M. Gotham, and M. Giraud, “Not all roads lead to rome: Pitch representation and model archi- tecture for automatic harmonic analysis,”Transactions of the International Society for Music Information Re- trieval (TISMIR), vol. 3, no. 1, pp. 42–54, 2020

    2020

  18. [18]

    Attend to chords: Improving harmonic analysis of symbolic music using transformer- based models,

    T.-C. Chen and L. Su, “Attend to chords: Improving harmonic analysis of symbolic music using transformer- based models,”Transactions of the International Soci- ety for Music Information Retrieval (TISMIR), vol. 4, no. 1, pp. 1–13, 2021

    2021

  19. [19]

    A deep learning method for enforcing coherence in automatic chord recognition,

    G. Micchi, K. Kosta, G. Medeot, and P. Chanquion, “A deep learning method for enforcing coherence in automatic chord recognition,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), Online, 2021, pp. 443–451

    2021

  20. [20]

    The neural autoregressive distribution estimator,

    H. Larochelle and I. Murray, “The neural autoregressive distribution estimator,” inProc. of the 14th Int. Conf. on Artificial Intelligence and Statistics (AISTATS), ser. Proceedings of Machine Learning Research, vol. 15, 2011, pp. 29–37

    2011

  21. [21]

    Interpretable music harmonic analysis through multi- linear mixture of experts,

    T. Triantafyllou, M. A. Nicolaou, and Y . Panagakis, “Interpretable music harmonic analysis through multi- linear mixture of experts,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

    2026

  22. [22]

    When in rome, listen to your corpus: A meta-corpus of functional harmony,

    M. Gotham, G. Micchi, N. N. López, and M. Sailor, “When in rome, listen to your corpus: A meta-corpus of functional harmony,”Transactions of the International Society for Music Information Retrieval (TISMIR), vol. 6, no. 1, pp. 150–166, 2023

    2023

  23. [23]

    The Annotated Mozart Sonatas: Score, Harmony, and Cadence,

    J. Hentschel, M. Neuwirth, and M. Rohrmeier, “The Annotated Mozart Sonatas: Score, Harmony, and Cadence,”Transactions of the International Society for Music Information Retrieval (TISMIR), vol. 4, no. 1, pp. 67–80, 2021. [Online]. Available: http://transactions.ismir.net/articles/10.5334/tismir.63/

    work page doi:10.5334/tismir.63/ 2021

  24. [24]

    musicbert-large,

    E. Karystinaios, “musicbert-large,” https://huggingface. co/manoskary/musicbert-large, 2026, model check- point, accessed 13 March 2026

    2026

  25. [25]

    GraphMuse: A Library for Symbolic Music Graph Processing

    E. Karystinaios and G. Widmer, “GraphMuse: A Library for Symbolic Music Graph Processing.” in Proceedings of the 25th International Society for Music Information Retrieval Conference, ISMIR 2024, San Francisco, California, USA and Online, November 10-14, 2024., 2024, pp. 651–658. [Online]. Available: https://doi.org/10.5281/zenodo.14877421

    work page doi:10.5281/zenodo.14877421 2024

  26. [26]

    Gradient surgery for multi-task learning,

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn, “Gradient surgery for multi-task learning,” in Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 2020

    2020

  27. [27]

    Conflict- averse gradient descent for multi-task learning,

    B. Liu, X. Liu, X. Jin, P. Stone, and Q. Liu, “Conflict- averse gradient descent for multi-task learning,” inAd- vances in Neural Information Processing Systems 34 (NeurIPS 2021), 2021. Proceedings of the 27th ISMIR Conference, Abu Dhabi, UAE, November 08–12, 2026

    2021

  28. [28]

    Gradnorm: Gradient normalization for adap- tive loss balancing in deep multitask networks,

    Z. Chen, V . Badrinarayanan, C.-Y . Lee, and A. Rabi- novich, “Gradnorm: Gradient normalization for adap- tive loss balancing in deep multitask networks,” inPro- ceedings of the 35th International Conference on Ma- chine Learning, ser. Proceedings of Machine Learning Research, vol. 80, 2018, pp. 794–803

    2018