Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

ARIMA: Reconstruction-Grounded Predictive Representation Learning for Symbolic Music

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A small window-level model learns symbolic-music features that match much larger token models on harmony, timing, and retrieval.

desk verdict Solid window-level SSL for piano MIDI that earns its keep on harmony/timing/retrieval; the efficiency claim is real but not cleanly isolated from the token-model control. read the letter →

arxiv 2607.10003 v1 pith:P4LG6Z5P submitted 2026-07-10 cs.SD

Mingyang Yao , Zhaoxiang Feng This is my paper · ORCID
classification cs.SD
keywords symbolicmusicself-supervisedlearningwindowrepresentationslatentpredictivearchitecturecontrastivenext-latentpredictionpianorollreconstructioninformationretrieval
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

Most self-supervised models for piano MIDI still train on tokenizer-specific event sequences, so time-span embeddings are only obtained by pooling and musical timing stays implicit. ARIMA instead cuts each performance into fixed-duration windows, compresses every window into a continuous latent that must reconstruct onset, sustain, chroma, and velocity, and trains a causal predictor with contrastive next-latent prediction so the latents also capture how music progresses. The resulting encoder and predictor embeddings, from a 38-million-parameter model, are especially strong on key estimation, tempo-ratio regression, performer verification, and cross-performance retrieval, and stay competitive with larger token baselines on style and emotion probes. Ablations show that next-latent prediction is what organizes local windows into temporally integrated features, while the structured reconstruction already prevents collapse without extra variance regularizers. A sympathetic reader cares because the design offers compact, analysis-ready time-span representations that keep low-level musical detail while modeling longer-range structure.

What carries the argument

ARIMA: each fixed-duration window is encoded to a continuous content latent that is grounded by structured reconstruction of onset, sustain, chroma and velocity rolls, while a causal predictor is trained by InfoNCE next-latent prediction so the same latents also model temporal progression.

What would settle it

Train a pure non-reconstructive next-latent baseline that converges on the same piano corpus and windowing; if it matches or beats ARIMA on key, IOI, verification and retrieval without the reconstruction heads, the reconstruction-grounding claim fails.

Watch

Extended reading notes

Core claim

ARIMA shows that reconstruction-grounded window latents plus contrastive next-latent prediction produce compact symbolic-music representations that are particularly effective for harmonic, timing, and cross-performance tasks and remain competitive with substantially larger token-level baselines on other understanding probes, with next-latent prediction essential for temporally integrated features.

Load-bearing premise

The design assumes that pure next-latent regression is unstable for multi-choice musical windows, and that the chosen reconstruction targets plus contrastive negatives already supply enough anti-collapse pressure and musical bias.

Editorial extensions

If this is right

  • Time-span music analysis can use fixed-duration continuous latents instead of tokenizer-tied token sequences.
  • Encoder latents and predictor states give complementary local-content versus temporally-integrated features for different MIR probes.
  • Structured musical reconstruction can replace explicit variance regularizers when learning predictive latents for symbolic music.
  • A 38M window model can match or exceed larger token baselines on harmony, timing and score-level retrieval tasks.

Reading between the lines

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

  • The same window-plus-predictor recipe could extend to multi-instrument or hierarchical (bar/phrase) symbolic SSL without changing the core objectives.
  • Because windows keep explicit timing and pitch, the latents may transfer more cleanly to editing or generation controls that need local musical detail than pure piece-level contrastive embeddings.
  • If next-window multi-modality is the real obstacle to pure regression, similar reconstruction grounding may help latent world models in other sparse, event-based domains.
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 / 6 minor

Summary. ARIMA is a self-supervised framework for symbolic piano music that learns fixed-duration window latents rather than tokenizer-specific token sequences. Each 2s window is encoded into a continuous content latent z_t supervised by structured reconstruction of onset/sustain pianorolls, chroma, and velocity; a causal Transformer predictor then produces temporal states h_t trained by contrastive next-latent prediction (InfoNCE with EMA targets and a carefully designed negative set). The dual representations are evaluated with linear/MLP probes on nine downstream tasks spanning composer/performer/emotion/key/difficulty classification, IOI regression, performer verification, and cross-performance retrieval. Against comparable-scale token baselines (including a same-corpus ~100M Aria replica) and larger public Aria models, ARIMA (38M) is strongest or second-strongest on several harmonic, timing, and retrieval tasks; ablations indicate that next-latent prediction is essential for temporally integrated features and that VICReg is unnecessary once reconstruction is present.

Significance. The paper addresses a genuine gap: most symbolic SSL yields time-span embeddings only by pooling token sequences whose length varies with density and tokenizer design. A compact window-level latent predictive model with explicit musical reconstruction is a coherent alternative, and the empirical package is stronger than typical MIR SSL papers—nine tasks, a same-corpus Aria replica, dual-representation analysis, and four ablations. Code release is promised. If the design advantages hold under tighter controls, ARIMA would be a useful inductive-bias template for analysis-oriented symbolic representation learning and a concrete adaptation of JEPA/CPC ideas to sparse event data. The significance is currently empirical and design-level rather than theoretical; it rests on how cleanly the comparisons isolate the proposed ingredients.

major comments (3)
  1. §4.4.1 and Table 1: the load-bearing efficiency claim (abstract; §5.1) that a 38M window-level model is “particularly efficient and effective” relative to token baselines, including the same-corpus Aria-100M replica, is not cleanly isolated. The replica remains an autoregressive token model (8 layers, d=768) under Aria’s objective and augmentation policy; it does not share ARIMA’s fixed-duration windows, reconstruction heads, EMA target encoder, InfoNCE negative rules, or dual z/h readout. Table 1 therefore confounds paradigm (window + reconstruction + next-latent) with tokenizer/event-sequence bias, pooling, and capacity allocation. A closer control—e.g., a window encoder of similar size trained only with reconstruction, or a token model given analogous reconstruction/prediction heads—would be needed before the strongest claim can be attributed to the proposed design rather than archite
  2. §5.3.2: the central design justification that pure non-reconstructive next-latent regression is unstable for symbolic music (multi-modal next windows; sparse low-level events) is supported only by unreported preliminary runs that “diverged or collapsed.” Because this premise motivates the reconstruction-grounded hybrid over a pure JEPA-style objective, the manuscript should either (i) report those failed configurations with enough detail to be reproducible, or (ii) add a stable non-reconstructive or alternative-reconstruction ablation that can be evaluated on the same probes. Without that evidence, the necessity of the chosen onset/sustain/chroma/velocity heads remains an assumption rather than a demonstrated result.
  3. §3.3.1 and §5.1: several of ARIMA’s strongest results (key estimation for z; IOI regression for h; cross-performance retrieval) align closely with features that are either reconstructed (chroma) or explicitly present in the note encoding (IOI, onset/duration). The paper notes inductive-bias alignment but does not quantify how much of the gain is reconstruction/encoding leakage versus learned temporal abstraction. A control that removes chroma reconstruction for the key probe, or that ablates the IOI input feature for the IOI probe, would clarify whether the claimed “temporally integrated representations” are doing the work on those tasks.
minor comments (6)
  1. Abstract and §1: the acronym expansion “Autoregressive Representation learning framework for Symbolic Music Analysis” does not match the title’s “ARIMA” or the method’s contrastive (not strictly autoregressive generative) objective; align naming.
  2. Eq. (1) and surrounding text: “relatjive” is a typo; also clarify units and clipping for notes that exceed K=100 per window.
  3. Figure 1 caption and §3.2: the figure labels “Content Encoder (online)” / “Target Encoder (stop-grad)” clearly, but the main text could state earlier that only z (not h) is reconstructed, to avoid confusion about which latent is grounded.
  4. Table 1: Aria-base (659M) scores are strikingly low on most probes relative to the 100M replica and Aria-embedding; a brief note on whether mean-pooling or the public checkpoint protocol is suboptimal would help readers interpret the foundation-scale rows.
  5. §4.1: training corpus size is given as “~15,000 piano performances” without total hours or note counts; adding a rough duration would make scale comparisons with Aria’s ~60k hours more transparent.
  6. §3.3.1: the double-weighting of same-piece different-window negatives is important; state the exact multiplier and whether it is applied after or before temperature scaling.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical SSL architecture paper whose claims rest on external downstream probes and component ablations, not on self-defining equations or load-bearing self-citations.

full rationale

ARIMA is a standard empirical representation-learning paper. The method defines fixed-duration window encoding, structured reconstruction heads (onset/sustain/chroma/velocity), an EMA target encoder, and InfoNCE next-latent prediction (Eqs. 2–9, loss (5)); these are design choices, not derivations that algebraically force the reported metrics. Downstream numbers (Table 1) are macro-F1, Spearman ρ, R², AUC and R@1 obtained by training linear/MLP probes or computing cosine similarities on held-out labeled sets (ATEPP, EMOPIA, POP909, CIPI, ASAP, Pianist8). Ablations (Table 2) simply remove prediction, merge onset/sustain, or add VICReg and re-measure the same external scores; none redefine the target by construction. Hyper-parameters (window length, λ weights, temperature, etc.) are chosen by hand but do not appear inside any claimed “prediction” formula. The sole self-citation ([23] Yao et al.) supplies only an auxiliary key-annotation resource and is not invoked as a uniqueness theorem or uniqueness result that forbids alternatives. Unreported pure-JEPA failures are narrative motivation, not a circular step. Consequently the derivation chain never reduces a claimed result to its own inputs; score 0 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

Load-bearing content is architectural and empirical rather than axiomatic derivation. The claim rests on standard transformer/InfoNCE mathematics, domain assumptions about what musical detail must be preserved, and a large set of hand-chosen hyperparameters that define the windowing and loss balance. No new physical entities are postulated; the dual latents z and h are architectural constructs whose value is measured by downstream probes.

free parameters (5)
  • window duration Δ and max notes K = Δ=2s, K=100
    Fixed at 2 s and 100 notes; defines the entire input granularity and therefore what ‘local content’ means. Not derived; chosen by design.
  • loss weights λ_pred, λ_recon, λ_vel and InfoNCE temperature = 0.25 / 1.0 / 5.0 / 0.07
    Balance reconstruction vs prediction and contrastive sharpness; warmed up and set by hand (λ_pred=0.25, λ_recon=1.0, λ_vel=5.0, τ_I=0.07). Directly control which features the latent is pressured to keep.
  • OWBCE kernel half-width and α, BCE positive weight ρ = half-width=3, α=12, ρ=3.0
    Control soft temporal margin around onsets and class imbalance; α=12, half-width 3, ρ=3.0 chosen without sweep reported.
  • EMA momentum τ and predictor context T = τ=0.996, T=60
    Target-encoder stability and effective history length (τ=0.996, T=60 windows ≈120 s); standard but still free design knobs that shape h.
  • model width/depth (d=512, 4-layer encoder, 6-layer predictor) = d=512, 4+6 layers, ~38M
    Capacity and inductive bias of both modules; fixed to yield ~38M parameters for the efficiency claim.
assumptions (4)
  • standard math Standard transformer attention, sinusoidal encodings, InfoNCE, and EMA target-encoder mathematics hold as used in CPC/BYOL/JEPA literature.
    Invoked throughout §3 for encoder, predictor, and contrastive loss; not re-proved.
  • domain assumption Low-level symbolic attributes (pitch, onset, sustain, velocity, chroma) carry direct analytical meaning that a pure non-reconstructive latent predictor will under-preserve.
    Stated in §1 and §5.3 as the reason reconstruction is added; motivates the whole reconstruction-grounded design.
  • domain assumption Fixed-duration windows with boundary-tied note flags lose no musical events and form a valid state sequence for next-latent prediction.
    §3.1.1 partitioning rule; if false, both reconstruction targets and temporal prediction are misaligned with the score.
  • ad hoc to paper In-batch negatives plus pitch-transposed same-window exclusion and same-piece hard negatives correctly encourage transposition invariance without piece fingerprinting.
    Negative-set construction in §3.3.1; specific to this training recipe and not independently validated outside the reported probes.
invented entities (2)
  • ARIMA dual latent (content z_t from window encoder + temporal h_t from causal predictor)
    purpose: Separate local musical content from causally integrated structure so different downstream tasks can use the appropriate representation.
    Core architectural product of the paper; value is measured only by the paper’s own probe suite, not by an external physical prediction.
  • Onset-weighted binary cross-entropy (OWBCE) with sine-window soft margin on binary onset targets
    purpose: Absorb quantization jitter of real-valued onsets onto a fixed grid without softening the target itself.
    Variant of soft-onset ideas from piano transcription; specific weighting formula (Eq. 8) is paper-defined.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ARIMA: Reconstruction-Grounded Predictive Representation Learning for Symbolic Music." pith.science (2026). https://pith.science/paper/P4LG6Z5P

@misc{pith2026260710003,
  author       = {Pith},
  title        = {Pith review of: ARIMA: Reconstruction-Grounded Predictive Representation Learning for Symbolic Music},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4LG6Z5P}},
  note         = {Machine review of arXiv:2607.10003}
}
read the original abstract

Self-supervised learning for symbolic music has advanced largely through token-level pretraining, but such representations remain tied to tokenizer-specific sequences and often provide time-span-level embeddings only indirectly. In this paper, we propose ARIMA, a reconstruction-grounded latent predictive framework for symbolic music that learns compact window-based representations directly from data. ARIMA encodes each fixed-duration window into a continuous latent representation, trains a causal predictor with contrastive next-latent prediction, and grounds the encoder through structured reconstruction of music elements. This design preserves local musical details while modeling temporal progression across windows. We evaluate ARIMA on downstream tasks spanning various levels of music understanding. Results show that ARIMA is particularly efficient and effective on tasks involving harmonic, timing, and cross-performance retrieval, while remaining competitive with much larger baselines on other tasks. Ablations further show that next-latent prediction is essential for temporally integrated representations, and that structured reconstruction stabilizes latent learning without requiring explicit variance regularization. The code is at https://github.com/AndyWeasley2004/symbolic_music_wm.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Helping Music Co-Creation Agents 'Listen' Well: Hierarchical Self-Supervised World Models for Understanding and Generation

    cs.SD 2026-08 conditional novelty 6.0 of 10

    A hierarchical self-supervised encoder on piano-roll images learns musical structure at different time scales, and a flow-matching decoder generates variations and inpaints gaps on CPU.

Reference graph

Works this paper leans on

34 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    INTRODUCTION Self-supervised learning (SSL) has become an important approach for learning transferable representations from un- labeled symbolic music. Recent works such as MidiBERT- Piano [1], MusicBERT [2], PianoBART [3], and Aria [4] adapt masked language modeling or autoregressive ob- jectives to MIDI tokenizations [5], including REMI [6], CP [7], Oct...

    arXiv 2026

  2. [2]

    We therefore re- view each area

    RELATED WORK Our work intersects with SSL for symbolic music and joint-embedding predictive frameworks. We therefore re- view each area. 2.1 Self-supervised learning for symbolic music. Masked language modeling (MLM) is the dominant pre- training paradigm for symbolic music. MidiBERT- Piano [1] adapts BERT-style pretraining and explores the effects of CP ...

  3. [3]

    We in- troduce the input representation, model components, and training objectives in the specification below

    METHOD Figure 1 illustrates the preprocessing pipeline, model back- bone, and main objectives in different modules. We in- troduce the input representation, model components, and training objectives in the specification below. 3.1 Input Representation 3.1.1 Window partitioning and note encoding. Given a piece of symbolic music, we partition it into a se- ...

  4. [4]

    Both used= 512, 8 heads, and an FFN dimension of 2048

    EXPERIMENTS 4.1 Data and Model Setting The window encoder is a 4-layer Transformer encoder, and the predictor is a 6-layer causal Transformer decoder with learned relative position bias. Both used= 512, 8 heads, and an FFN dimension of 2048. Training is regularized with dropout and stochastic depth [17] probability of 0.1. Each window spans∆ = 2s with at ...

  5. [5]

    next-window

    RESULTS 5.1 Main Results 5.1.1 Comparable-scale baselines. Table 1 summarizes performance across all tasks. Among comparable-scale models, ARIMA is consistently compet- itive despite having the smallest parameter count. The pre- dictor representationhobtains the best results on com- poser classification, inter-onset interval regression, and performer veri...

  6. [6]

    CONCLUSION In this paper, we propose ARIMA, a reconstruction- grounded latent predictive framework for symbolic music representation learning. ARIMA encodes fixed-duration music windows into compact continuous latents, grounds these latents through structured reconstruction, and models temporal progression with contrastive next-latent predic- tion. Experi...

  7. [7]

    All scientific decisions, experimental de- signs, figure drawing, results analysis and discussion were completely made by humans

    ACKNOWLEDGEMENT Generative AI tools assisted with manuscript revision and proofreading. All scientific decisions, experimental de- signs, figure drawing, results analysis and discussion were completely made by humans

  8. [8]

    MidiBERT-Piano: Large-scale pre-training for symbolic music classification tasks,

    Y .-H. Chou, I.-C. Chen, J. Ching, C.-J. Chang, and Y .- H. Yang, “MidiBERT-Piano: Large-scale pre-training for symbolic music classification tasks,”Journal of Creative Music Systems, vol. 8, no. 1, 2024

    2024

Show all 34 references
  1. [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 Asso- ciation for Computational Linguistics: ACL-IJCNLP 2021, 2021, pp. 791–800

    2021

  2. [10]

    PianoBART: Symbolic piano music genera- tion and understanding with large-scale pre-training,

    X. Liang, Z. Zhao, W. Zeng, Y . He, F. He, Y . Wang, and C. Gao, “PianoBART: Symbolic piano music genera- tion and understanding with large-scale pre-training,” inProc. of the IEEE Int. Conf. on Multimedia and Expo (ICME), 2024

    2024

  3. [11]

    Scaling self-supervised representation learning for symbolic piano performance,

    L. Bradshaw, H. Fan, A. Spangher, S. Biderman, and S. Colton, “Scaling self-supervised representation learning for symbolic piano performance,” inProc. of the 26th Int. Society for Music Information Retrieval Conf., 2025

    2025

  4. [12]

    Natural language processing methods for symbolic music generation and information retrieval: A survey,

    D.-V .-T. Le, L. Bigo, D. Herremans, and M. Keller, “Natural language processing methods for symbolic music generation and information retrieval: A survey,” ACM Computing Surveys, vol. 57, no. 7, pp. 1–40, 2025

    2025

  5. [13]

    Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions,

    Y .-S. Huang and Y .-H. Yang, “Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions,” inProc. of the 28th ACM Int. Conf. on Multimedia, 2020, pp. 1180–1188

    2020

  6. [14]

    Compound word transformer: Learning to compose full-song music over dynamic directed hypergraphs,

    W.-Y . Hsiao, J.-Y . Liu, Y .-C. Yeh, and Y .-H. Yang, “Compound word transformer: Learning to compose full-song music over dynamic directed hypergraphs,” inProc. of the AAAI Conf. on Artificial Intelligence, vol. 35, no. 1, 2021, pp. 178–186

    2021

  7. [15]

    Clamp: Contrastive language-music pre-training for cross- modal symbolic music information retrieval,

    S. Wu, D. Yu, X. Tan, and M. Sun, “Clamp: Contrastive language-music pre-training for cross- modal symbolic music information retrieval,” in Proceedings of the 24th International Society for Music Information Retrieval Conference, ISMIR 2023, 2023, pp. 157–165. [Online]. Avail...

    2023 doi

  8. [16]

    Self-supervised learning from images with a joint- embedding predictive architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Ballas, “Self-supervised learning from images with a joint- embedding predictive architecture,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 15 ...

    2023

  9. [17]

    Revisiting fea- ture prediction for learning visual representations from video,

    A. Bardes, Q. Garrido, J. Ponce, X. Chen, M. Rabbat, Y . LeCun, M. Assran, and N. Ballas, “Revisiting fea- ture prediction for learning visual representations from video,”arXiv preprint arXiv:2404.08471, 2024

    2024 arXiv

  10. [18]

    Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels,

    L. Maes, Q. L. Lidec, D. Scieur, Y . LeCun, and R. Balestriero, “Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels,” arXiv preprint arXiv:2603.19312, 2026

    2026 arXiv

  11. [19]

    Representa- tion learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representa- tion learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  12. [20]

    Improving BERT for symbolic music understanding using token denoising and pi- anoroll prediction,

    J. Wang and L. Su, “Improving BERT for symbolic music understanding using token denoising and pi- anoroll prediction,” inProceedings of the 26th Inter- national Society for Music Information Retrieval Con- ference, ISMIR 2025, J. Nam, D. Jeong, K. Choi, L. Su, M. Fuentes, T. N...

    2025

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  14. [22]

    Bootstrap your own la- tent – a new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko, “Bootstrap your own la- tent – a new approach to self-supervised learning,” in Advances in Neural Info...

    2020

  15. [23]

    High- resolution piano transcription with pedals by regress- ing onset and offset times,

    Q. Kong, B. Li, X. Song, Y . Wan, and Y . Wang, “High- resolution piano transcription with pedals by regress- ing onset and offset times,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3707–3717, 2021

    2021

  16. [24]

    Deep networks with stochastic depth,

    G. Huang, Y . Sun, Z. Liu, D. Sedra, and K. Q. Wein- berger, “Deep networks with stochastic depth,” inProc. of the European Conf. on Computer Vision (ECCV), 2016, pp. 646–661

    2016

  17. [25]

    ATEPP: A dataset of automatically tran- scribed expressive piano performance,

    H. Zhang, J. Tang, S. R. M. Rafee, S. Dixon, and G. Fazekas, “ATEPP: A dataset of automatically tran- scribed expressive piano performance,” inProc. of the 23rd Int. Society for Music Information Retrieval Conf., 2022

    2022

  18. [26]

    POP909: A pop-song dataset for music arrangement generation,

    Z. Wang, K. Chen, J. Jiang, Y . Zhang, M. Xu, S. Dai, and G. Xia, “POP909: A pop-song dataset for music arrangement generation,” inProc. of the 21st Int. Soci- ety for Music Information Retrieval Conf., 2020

    2020

  19. [27]

    EMOPIA: A multi-modal pop piano dataset for emotion recognition and emotion-based music gen- eration,

    H.-T. Hung, J. Ching, S. Doh, N. Kim, J. Nam, and Y .- H. Yang, “EMOPIA: A multi-modal pop piano dataset for emotion recognition and emotion-based music gen- eration,” inProc. of the 22nd Int. Society for Music In- formation Retrieval Conf., 2021, pp. 318–325

    2021

  20. [28]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. of the Int. Conf. on Learning Representations (ICLR), 2019

    2019

  21. [29]

    Pianist8 dataset,

    Y .-H. Chou, I.-C. Chen, C.-J. Chang, J. Ching, and Y .- H. Yang, “Pianist8 dataset,” Zenodo, 2021

    2021

  22. [30]

    Bachi: Boundary-aware symbolic chord recognition through masked iterative decoding on pop and classi- cal music,

    M. Yao, K. Chen, S. Dubnov, and T. Berg-Kirkpatrick, “Bachi: Boundary-aware symbolic chord recognition through masked iterative decoding on pop and classi- cal music,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE,...

    2026

  23. [31]

    Combining piano perfor- mance dimensions for score difficulty classification,

    P. Ramoneda, D. Jeong, V . Eremenko, N. C. Tamer, M. Miron, and X. Serra, “Combining piano perfor- mance dimensions for score difficulty classification,” Expert Systems with Applications, vol. 238, p. 121776, 2024

    2024

  24. [32]

    ASAP: A dataset of aligned scores and performances for piano transcription,

    F. Foscarin, A. McLeod, P. Rigaux, F. Jacquemard, and M. Sakai, “ASAP: A dataset of aligned scores and performances for piano transcription,” inProc. of the 21st Int. Society for Music Information Retrieval Conf., 2020, pp. 534–541

    2020

  25. [33]

    VICReg: Variance-invariance-covariance regularization for self- supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “VICReg: Variance-invariance-covariance regularization for self- supervised learning,” inProc. of the Int. Conf. on Learning Representations (ICLR), 2022

    2022

  26. [34]

    Learning from reward-free offline data: A case for planning with latent dynamics mod- els,

    V . Sobal, W. Zhang, K. Cho, R. Balestriero, T. G. Rud- ner, and Y . LeCun, “Learning from reward-free offline data: A case for planning with latent dynamics mod- els,”arXiv preprint arXiv:2502.14819, 2025

    2025

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.