Martin Takac

Mohamed bin Zayed University of Artificial Intelligence

Unclaimed · AI-structured profileOpenAlex

ML researcher at MBZUAI spanning optimal-transport generative modeling and adversarial robustness in vision

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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—

Not yet scored: Evolve holds only two 2026 arXiv preprints for this name, with no publication counts, venue tiers, or authorship-share metadata. This axis would measure total scholarly output and its venue distribution once a fuller bibliographic record is linked.

Impact—

Insufficient data: no citation counts, h-index, or venue-level influence signals exist in the current snapshot. This axis would measure how widely this researcher's publications are cited and adopted, which cannot be estimated from two recent arXiv preprints alone.

Focus0/100

The focus score is 0, consistent with a field distribution that splits exactly 50/50 between adversarial machine learning (computer vision) and machine learning (generative models / optimal transport) across just two reviewed papers — one per field. This reads as two parallel early-stage directions rather than concentrated specialization, and both entries are 2026 arXiv preprints, so the long-term focus may not have settled yet.

Collaboration46/100

The collaboration score is 46/100, based on 9 distinct co-authors in the current record. The frequent-coauthor list spans both research tracks and includes Mohammed-Yassine Habibi, Klea Ziu, Makoto Yamada, Semyon Semenov, Viktor Kovalchuk, Meir Roketlishvili, Albert Baichorov, Fakhri Karray, and Arip Asadulaev. On both catalog papers Takac appears in a middle author position without corresponding-author status, indicating team-based rather than solo or lead-corresponding work so far.

Activity—

Not computable yet: with only two 2026 preprints and no year-by-year publication record, recent output velocity cannot be scored. This axis would track publication cadence and sustained research activity across years once a complete bibliographic timeline is available.

Reproducibility38/100

The reproducibility score averages 38 across the 2 papers Evolve has reviewed. Branched Optimal Transport Amortization is rated low readiness, consistent with proof-of-concept evidence and modest, mostly qualitative gains; the oscillatory predictive-learning robustness paper is rated medium, though the review flags CIFAR-only evaluation, a narrow robust regime, and single-seed ablations as reproducibility risks. Reproducibility is therefore early and mixed rather than established.

Research summary

Martin Takac is a researcher at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) whose reviewed work spans two distinct machine-learning directions: generative modeling via optimal transport, and adversarial robustness in computer vision. In the optimal-transport track, the 2026 arXiv preprint 'Branched Optimal Transport Amortization' adds a concave economy-of-scale branched-OT cost to flow matching; Evolve's review judges the concept sound and promising, but the evidence proof-of-concept with modest, mostly qualitative gains. In the robustness track, 'Neither Adversarial Training Nor Purification' reports emergent adversarial robustness from oscillatory predictive learning built on Kuramoto-oscillator (AKOrN) architectures, claiming 76.63% CIFAR-10 robust accuracy without adversarial training — compelling, though CIFAR-only, with a narrow robust regime and single-seed ablations. The bibliometric field distribution splits exactly 50/50 between these two fields, and the current collaboration network contains 9 distinct co-authors, including Mohammed-Yassine Habibi, Makoto Yamada, and Fakhri Karray. Output, impact, and activity axes are not yet scored because citation and venue metadata are missing; the profile is correspondingly early-stage and rests on two arXiv preprints.

Collaboration pattern

Current data records 9 distinct co-authors, with a collaboration score of 46/100. The frequent-coauthor list includes Mohammed-Yassine Habibi, Klea Ziu, Makoto Yamada, Semyon Semenov, Viktor Kovalchuk, Meir Roketlishvili, Albert Baichorov, Fakhri Karray, and Arip Asadulaev — a pool spanning both the optimal-transport and adversarial-robustness tracks. On both catalog papers Takac sits in a middle author position without a corresponding-author flag, so the visible pattern is team-based collaboration rather than solo or corresponding-lead work.

Research keywords

Branched optimal transportFlow matchingBenamou–Brenier dynamic formulationOptimal transport amortizationContinuous normalizing flowsGenerative modelingAdversarial robustnessKuramoto oscillator networksSelf-supervised learningPredictive learningAdversarial machine learning

Highlights

Branched optimal transport cost for flow matching
The 2026 preprint adds a concave economy-of-scale branched-OT cost to flow matching, extending Benamou–Brenier-style dynamic optimal transport toward amortized generative modeling with continuous normalizing flows. Evolve judges the concept sound and promising while noting the evidence remains proof-of-concept, with modest, mostly qualitative gains.
76.63% CIFAR-10 robust accuracy without adversarial training
The oscillatory predictive-learning paper (AKOrN / Kuramoto oscillators, self-supervised learning, X-PhiNet) reports emergent adversarial robustness, claiming 76.63% CIFAR-10 robust accuracy with neither adversarial training nor purification. Evolve finds the claim compelling but bounded: CIFAR-only evaluation, a narrow robust regime, and single-seed ablations.
A two-front 2026 research agenda
Both catalog entries appeared in 2026 on arXiv and split evenly between the two fields — generative modeling via optimal transport and adversarial robustness in computer vision — an exact 50/50 field distribution that signals two parallel directions at an early stage rather than a settled specialty.

Timeline

2026Posts 'Branched Optimal Transport Amortization' to arXiv, adding a concave economy-of-scale branched-OT cost to flow matching for amortized generative modeling.
2026Posts the oscillatory predictive-learning robustness paper, claiming 76.63% CIFAR-10 robust accuracy without adversarial training or purification.
2026Both preprints enter the Evolve review catalog, with reproducibility readiness rated low and medium respectively, marking the start of this early-stage profile.

Papers in the Evolve catalog

Frequent co-authors

Mohammed-Yassine Habibi×1 Klea Ziu×1 Makoto Yamada×1 Semyon Semenov×1 Viktor Kovalchuk×1 Meir Roketlishvili×1 Albert Baichorov×1 Fakhri Karray×1 Arip Asadulaev×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.

Bibliometric scores are derived from OpenAlex-based bibliometric data only; citation counts, venue tiers, and yearly output are absent, so the output, impact, and activity axes are unscored.
The profile rests on just two 2026 arXiv preprints reviewed by Evolve; it is an early-stage profile and may not reflect the researcher's full publication record.
No ORCID is provided; 'Martin Takac' is name-resolved from public bibliographic data and may conflate distinct researchers sharing the name.
The reproducibility score (mean 38) is computed from only 2 reviewed papers, so it is indicative rather than stable.