Marcel van Gerven

Radboud University Nijmegen · Netherlands

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher at Radboud University working on spatial audio ML and unsupervised anomaly detection.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-09) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

The output axis is unscored: the snapshot contains no reliable career publication total — h-index and i10-index are both 0 and only 3 works are recorded — so overall output volume cannot yet be measured from this data.

Impact—

Impact cannot be scored: no citation counts, h-index trajectory, or field-weighted influence metrics are available in this snapshot. The axis would measure how widely and how strongly other researchers cite and build on this work.

Focus0/100

Focus scores 0/100 because the field distribution is perfectly flat: spatial-audio ML, unsupervised anomaly detection, and SELD each take 33.3% of the record. OpenAlex topics reinforce a coherent but spread audio-ML cluster — Speech Recognition and Synthesis, Speech and Audio Processing, and Music and Audio Processing (2 works each), plus Anomaly Detection and Adversarial Robustness (1 each). The work is thematically adjacent rather than dispersed, but no single field yet dominates.

Collaboration39/100

Collaboration scores 39/100. The recent window shows 6 distinct coauthors across 3 works, with recurring partners including Mehmet Yamaç, Lei Xu, Goksenin Yuksel, Kiki van der Heijden, Yagmur Mustu, and Muhammad Numan Yousaf. Authorship positions in the catalog are middle or last, with no corresponding-author role, suggesting a contributing senior-collaborator pattern within small ML teams rather than a solo or heavily networked profile.

Activity40/100

Activity scores 40/100 on thin recent volume: 3 works in the recent three-year window, all captured under 2026 and all arXiv preprints. The three papers form one concentrated burst — EquiSELD, Bearings, and the PoS anomaly-detection theory paper — rather than a sustained multi-year cadence, so activity should be re-read once the OpenAlex record accumulates more publication years.

Reproducibility85/100

Reproducibility is the strongest axis at 85/100, computed across all 3 catalog papers (reproMean 85). EquiSELD and Bearings are rated high reproducibility readiness — both report concrete ablations such as the O(3)-versus-SO(3) comparison and the F20° <4→50.4 TAU gain — while the PoS anomaly-detection paper is medium, with its idealized geometric assumptions flagged as contested. Overall an above-average, evidence-heavy reporting style.

Research summary

Marcel van Gerven is a machine learning researcher at Radboud University Nijmegen (NL) whose Evolve catalog centers on two adjacent strands: machine learning for spatial audio and unsupervised anomaly detection. The bibliometric snapshot is OpenAlex name-resolved and unusually thin — h-index 0, i10-index 0, and 3 recent works, all dated 2026 — so this should be read as an early-stage profile of a research line rather than a career-level assessment. Within that line, the 2026 output is coherent: EquiSELD trains exactly O(3)-equivariant attention networks for sound event localization and detection, winning on TAU2021 while its O(3)-versus-SO(3) ablation shows no clear gain; Bearings introduces self-supervised soundfield embeddings from first-order Ambisonics that plug into frozen encoders, lifting ablated SELD F20° from under 4 to 50.4 on TAU; and a last-author theory paper uses Pursuit-of-Subspaces geometry to explain when reconstruction-based anomaly detection succeeds or fails, proposing Dynamic Push–Pull and nested carving. Field distribution splits evenly (~33% each) across spatial-audio ML, unsupervised anomaly detection, and SELD. Reproducibility is the strongest dimension (85/100, all 3 papers assessed, two rated high readiness); collaboration (39) and activity (40) are modest, with 6 distinct coauthors.

Collaboration pattern

Van Gerven's catalog work is produced in small machine-learning teams: 6 distinct coauthors appear across the 3 recent papers, with recurring partners Mehmet Yamaç, Lei Xu, Goksenin Yuksel, Kiki van der Heijden, Yagmur Mustu, and Muhammad Numan Yousaf. His listed positions are middle or last author and none of the three papers names him as corresponding author, consistent with a senior technical contributor who anchors theory and methodology rather than leading every submission.

Research keywords

sound event localization and detectionfirst-order Ambisonicsself-supervised learningsoundfield embeddingsequivariant neural networksO(3) equivariancereconstruction-based anomaly detectionautoencodersunsupervised representation learningmasked autoencoderspatial audioadversarial robustness

Highlights

Exactly O(3)-equivariant sound event localization and detection
EquiSELD trains an exactly O(3)-equivariant attention network for SELD and leads the TAU2021 benchmark, with high reproducibility readiness. Evolve's caveat: the O(3)-versus-SO(3) ablation shows no clear gain and the STARSS23 wins fall within noise, so exact equivariance is efficient but not yet decisively better.
Self-supervised soundfield embeddings that plug into frozen encoders
Bearings claims the first self-supervised soundfield embeddings from first-order Ambisonics that can be plugged into frozen encoders. Ablated SELD gains are solid — F20° rises from under 4 to 50.4 on TAU — though the lead over hand-crafted statistics is moderate. Rated high reproducibility readiness.
A geometric account of when reconstruction-based anomaly detection fails
A last-author theory paper argues that Pursuit of Subspaces (PoS) geometry explains when reconstruction-based anomaly detection succeeds or fails, and proposes Dynamic Push–Pull and nested carving to keep improving detection. Evolve notes its idealized assumptions remain contested at the margin; reproducibility readiness is medium.

Timeline

2026Posted EquiSELD on arXiv: efficient training of exactly O(3)-equivariant sound event localization and detection networks, leading on TAU2021.
2026Released Bearings on arXiv: self-supervised soundfield embeddings from first-order Ambisonics that plug into frozen encoders.
2026Co-authored the PoS-perspective theory paper on when reconstruction-based anomaly detection succeeds or fails, as last author.

Papers in the Evolve catalog

Frequent co-authors

Goksenin Yuksel×2 Kiki van der Heijden×2 Mehmet Yamaç×1 Yagmur Mustu×1 Muhammad Numan Yousaf×1 Lei Xu×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometrics come from an OpenAlex name-resolved identity only; no ORCID was provided, so the record may be incomplete or partially conflated with namesakes.
The snapshot shows h-index 0, i10-index 0, and only 3 works (all 2026), which points to an early or incomplete identity cluster rather than verified career totals.
Output and impact dimensions are unscored because citation counts and career publication totals are absent from the data.
All three catalog papers are 2026 arXiv preprints; peer-review status and venue acceptance are not captured.