acceptodds
Desk rejected at ICLR 2027

PianoFingerprints: Rediscovering Chopin via Composition-Aware Contrastive Learning

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Music discovery from symbolic archives requires relationships among works that persist across heterogeneous collections and MIDI realizations. We propose PianoFingerprints, a composition-anchored multilevel atlas combining co-registered three-channel time-pitch MIDI tensors, ViT-Base, and source-priority multi-positive contrastive learning. A segment-realization-work hierarchy separates reproducible work structure from archive and realization variation. Across 47,489 realizations from 11 archives, KDF is held out from training for cross-archive evaluation and fitted separately for single-archive discovery. Held-out KDF contains 232 nonempty realizations of 139 normalized works. The performance-archive geometry agrees with work geometry independently aggregated from held-out KDF (), showing that relative work similarities recur across archives. Two nonoverlapping halves of the KDF realizations also yield similar work geometries (), indicating stability across different realizations of the same works. In the single-archive atlas, mean within-genre similarity exceeds mean cross-genre similarity ( versus ). There is no evidence that work distinctiveness has a statistically significant global linear association with composition year (, ). For 66 cross-genre neighbour pairs reproduced in both archive views, the closest ten-second passages have a mean symbolic similarity of . Shuffling specific work identities within each genre, while retaining the same pair of genres for every relation, yields a permutation mean of (, ). Together, these results position PianoFingerprints as a corpus-conditioned discovery instrument for Chopin archives, using cross-archive replication to identify reproducible structure and local passages to support interpretation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
composition-aware contrastive learning, symbolic music, music representation learning, music discovery, music information retrieval, computational musicology

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jrxWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a technique for learning an embedding model for symbolic piano MIDI files, aimed at extracting multi-level compositional information. This primarily works by chunking MIDI files, representing them as a multi-channel 2D tensor, embedding them with a ViT, and training the model using an InfoNCE-style objective where positive pairs derive from the same composition. The authors mean-pool embeddings to produce 'realization', archive, and composition-level embeddings, which they analyze using cosine similarity. The authors find their approach performs competitively at tasks like identity retrieval, and reveal consistent relationships between archive-level embeddings judged by quantities like embedding cosine similarity.

Critical strengths

The authors explore a potentially interesting problem of learning composition-level representations using contrastive learning. Examination of cross-archive inter-piece relationships is an interesting methodology. The proposed approach performs the best in some evaluations when compared to the authors' choice of baselines.

Critical weaknesses

There are several areas which could use improvement.

Clarity:

In many places the document is quite hard to parse. The prose should be improved significantly before publication. A few examples:

title: 'Rediscovering Chopin'. What does this term mean?
l16: 'KDF' is a non-standard acronym used without being defined.
l18: 'Performance-archive geometry'. This term should be more clearly explained.
l30: 'corpus-conditioned discovery instrument for Chopin archives'. What is a discovery instrument? Does the proposed model just model Chopin's works?
l383: 'Realization split-half reliability evaluates reconstruction within held-out KDF'. This is very hard to parse.
l435: 'The KDF atlas therefore does not support a single monotonic account of Chopin’s compositional development'. Also hard to parse.

Methodological novelty:

The use of segment-level contrastive learning techniques to generate embeddings, and the exploration of compositional relationships in MIDI music have precedent [1], which the authors compare to in this work. There is also a long history of work on music fingerprinting, which is presumably applicable to piano performance audio. In light of this, the methodological novelty/contributions may then be framed as primarily (1) archival differences, (2) data sampling differences, i.e., sampling positive pairs from the different realizations/performances of the same composition, and (3) the evaluations of compositional relationships that the authors introduce. The reviewer views this as potentially below the bar for ICLR, and perhaps better suited for a music-oriented publication venue such as ISMIR.

Evaluations:

Given the methodological contributions, the missing experiment is a controlled comparison that isolates the proposed representation, backbone, sampling strategy. This could include, e.g., fine-tuning other contrastively trained MIDI models on the same data, or otherwise following its methodology, to isolate the effects of the data sampling and training procedure that is used (e.g., using positive pairs from same-composition segments). It would then be easier to evaluate the main contributions of the paper.

[1] Bradshaw, L., Fan, H., Spangher, A., Biderman, S., & Colton, S. (2025). Scaling Self-Supervised Representation Learning for Symbolic Piano Performance. https://arxiv.org/abs/2506.23869

Questions for authors

See weaknesses.

Reviewer_otrEOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Learns a contrastive embedding of Chopin piano MIDI (ViT on 3-channel piano rolls), aggregated from segments to realizations to works. Trains on 10 performance archives and tests on held-out KDF. Checks whether work-to-work similarity replicates across archives (r_s = 0.77) and across realization halves.

Critical strengths

Interesting goal: work-level structure that holds across heterogeneous MIDI sources. Sensible evaluation idea: cross-archive replication and split-half reliability rather than retrieval alone. Large assembled corpus, several baselines (matched objectives, pretrained symbolic models, handcrafted features). Clear figures, thorough appendix.

Critical weaknesses

Proposed objective doesn’t beat its own ablation. Single run per method, no seeds? Neighbour validation possibly circular: checking features resemble model inputs (pitch, density, velocity). Genre structure reported without baselines. Musical meaningfulness of discovered neighbours not assessed (key, tempo, texture confounds). Limited ML novelty for ICLR.

Questions for authors

How do genre and neighbour results compare against random-encoder or descriptor baselines? Are discovered cross-genre neighbours explained by key, tempo or texture alone? Any expert or musicological validation of the neighbours? Any overlap between KDF and the training archives?

Reviewer_GJYKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds a composition-anchored multilevel atlas for discovering relations among Chopin works across heterogeneous MIDI archives. Inputs are 10-second windows of a co-registered three-channel time-pitch tensor (note activity, normalized velocity, attack density) encoded by a ViT-Base and trained with batch-all supervised contrastive learning (Khosla et al. SupCon) plus a source-priority positive sampler that prefers same-work positives across archives, then realizations, then windows. Segment embeddings are aggregated hierarchically (segment, realization, archive, work levels) into source-balanced work fingerprints. Two regimes are studied: cross-archive (train on 10 performance-MIDI archives, 47,256 realizations, validate on held-out Kunst der Fuge) and single-archive (fit KDF directly). Reported results: cross-archive work-geometry agreement rs=0.767, split-half reliability rs=0.870, within-genre similarity (0.199) above cross-genre (0.045), no significant linear association between distinctiveness and composition year (r=-0.148, p=0.075), and cross-genre neighbours supported by local passages under a genre-preserving permutation test (delta=0.145, p=1e-4). Baselines include matched SupCon/NCE/InfoNCE/CE/triplet ViT controls and pretrained Aria, MusicBERT, CLaMP 2, plus a fixed symbolic descriptor.

Critical strengths
  1. The cross-archive replication protocol is the paper's real asset. Asking whether a 129x129 work-similarity matrix built from ten independently assembled archives recurs in a held-out archive is a rare and valuable validation discipline in computational musicology, and the permutation test grounding the local-passage support (0.548 vs 0.403, p=1e-4) is cleanly designed.

  2. The statistical care is consistent. Source-balanced aggregation, split-half reliability, and permutation-based significance tests are applied throughout, which is not common in a field prone to overinterpreting embeddings.

Critical weaknesses
  1. The ML contribution is marginal, and the paper's own mechanism does not improve its central result. The encoder is ViT-Base with standard SupCon, and the only novelty is the source-priority sampler. On the headline cross-archive geometry metric PianoFingerprints (rs=0.767) sits below its own matched SupCon/NCE/InfoNCE controls (rs=0.780 to 0.811), with its advantage confined to the secondary split-half reliability metric (0.870 vs 0.841). Standard SupCon also beats it on full-realization Recall@1 (0.846 vs 0.824). So the source-priority multi-positive learning contribution is hard to defend.

  2. The substantive findings are descriptive or null and rest on small samples. Several genres have 2 to 4 works, within-genre similarity (0.199) is modest, and the chronology result is a marginal null (p=0.075 reported as no evidence). The musicological conclusions (no single linear style trajectory, uneven genre organization) are largely established already.

Questions for authors
  1. Why does source-priority sampling fail to improve cross-archive geometry over Standard SupCon, and what is the mechanism by which it improves only split-half reliability? This is central to whether the method is a genuine advance or a re-parameterization of SupCon.

  2. How robust are the within-genre similarity claims given only 2 to 4 works for ballade, fantasie, impromptu, scherzo, and sonata? Would leave-one-work-out or bootstrap over works change the graded-genre-organization conclusion?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.