{"resourceId":"sydney-riddled-basins-dl-reproducibility-nmi-20261105","versions":[{"version":"external-97ab48e426c0c76ebe0f92936f37651d7b20da5060b04b1db63fd22dbb71a231","resource":{"id":"sydney-riddled-basins-dl-reproducibility-nmi-20261105","title":"Fractal basin geometry and the limits to predictability and reproducibility in deep learning","organization":"University of Sydney (School of Physics)","sector":"University research (fundamental limits of deep-learning training reproducibility)","geography":"Australia (international; transferable to any research university training ML models)","publishedAt":"5 November 2026 (Nature Machine Intelligence)","publicationDate":"2026-11-05","eventDate":null,"sourceName":"Nature Machine Intelligence","sourceLabel":"Ly, A. & Gong, P., \"Fractal basin geometry and the limits to predictability and reproducibility in deep learning,\" Nature Machine Intelligence, 5 Nov 2026, doi:10.1038/s42256-026-01336-3","sourceUrl":"https://www.nature.com/articles/s42256-026-01336-3","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["developers-agents","infrastructure","operating-model"],"finding":"University of Sydney researchers argue that deep-learning training outcomes can be governed by riddled (fat-fractal) basins of attraction: every initialization leading to one solution lies arbitrarily close to initializations leading to different solutions. In a minimal model they derive a near-zero uncertainty exponent (about 0.0126), implying that large increases in initialization precision yield only marginal predictability gains. In deep networks they report near-zero parameter-space exponents on both GPU and deterministic CPU runs, and a functional uncertainty exponent near zero with churn plateauing around 2.9%, so functional outcomes also resist improvement from finer initialization precision.","sledRelevance":"Relevant to research computing and AI-for-science groups that treat single training runs as authoritative results. The paper frames irreproducibility as a dynamical-systems limit rather than only seed or GPU nondeterminism, affecting how campuses should design multi-seed evaluation and model acceptance.","evidence":"Peer-reviewed Nature Machine Intelligence article; analytical minimal model plus empirical signatures across tested architectures, optimizers and protocols. Free abstract and substantial HTML body inspected; some figures/paywalled supplements not fully used. Authors Ly and Gong (University of Sydney) confirmed via journal metadata and institutional news.","architectureImplications":"Training pipelines for research models should assume outcome-level sensitivity that seeding alone may not remove; ensemble or multi-seed reporting and functional similarity checks become first-class infrastructure, not optional QA.","governanceImplications":"Research integrity and computational-reproducibility policies should distinguish controllable extrinsic noise from possible intrinsic riddling-like limits, and require documentation of seeds, hardware and functional variability.","securityPrivacyImplications":"Not a security study; implications are reliability and trustworthiness of deployed research models rather than confidentiality.","caveats":"Mechanism proposed as a candidate organizing principle; not every architecture or task proven riddled; practical mitigation recipes beyond diagnostics are limited in the inspected text; institutional press piece is dated 6 Nov and was used only to confirm affiliation.","streamIds":["research"],"roles":{"sales":"Reliability framing: Nature Machine Intelligence reports a University of Sydney result that deep-learning training can hit a scale-insensitive reproducibility limit (near-zero uncertainty exponents, including on CPU). That supports bounded conversations with research computing and AI-for-science leads about multi-seed evaluation budgets. Interpretation: the paper is theoretical/empirical analysis, not a product failure report, so it cannot support claims that any campus cluster is defective.","engineering":"Fit is ML training platforms and AI-for-science pipelines that publish single-run metrics. Architecture should capture seeds, hardware, multi-seed ensembles and functional churn metrics. Prerequisites include deterministic CPU baselines where feasible and storage for replicate runs. Constraints: riddling is a proposed mechanism; not a turnkey fix. Proof of value: measure functional churn across epsilon-close initializations on a campus benchmark model.","delivery":"Delivery work is a reproducibility diagnostic for one high-stakes research model family: replicate training with controlled initialization perturbations, report parameter and functional uncertainty exponents or churn, and update acceptance criteria. Ownership sits with research computing and the PI. Governance checkpoints include documenting seeds/hardware and requiring multi-seed evidence for claims. Acceptance criteria are proposals such as published churn bounds; risks include over-interpreting one paper as blocking all deep learning."},"roleDetails":{"sales":{"customerProblem":"Research groups often treat one training run as definitive, then struggle when peers cannot reproduce model behavior.","stakeholders":["Research computing director","AI-for-science PIs","Research integrity officer","Graduate program leads"],"discoveryQuestions":["How many seeds are required before a campus result is publishable?","Do you measure functional differences across near-identical runs?","Is GPU nondeterminism documented in methods?","Which models are high-stakes enough for ensemble acceptance?"],"valueHypothesis":"Budgeting multi-seed and functional-churn checks may reduce irreproducible campus AI claims even when finer initialization precision does not help.","potentialEngagement":"A reproducibility diagnostic engagement on one production research model family.","unsupportedClaims":["Campus GPUs are uniquely broken","Deep learning cannot be used in research","Seeding alone guarantees reproducibility"]},"engineering":{"fit":"Strong for shared ML platforms that train research models; weaker for inference-only services.","architectureIntegration":"Training job metadata (seed, hardware, dtype), multi-seed orchestration, functional similarity/churn metrics, optional CPU deterministic baselines.","prerequisites":["Job provenance logging","Capacity for replicate runs","Agreement on functional similarity metrics"],"constraints":["Riddling is a candidate mechanism","Compute cost of ensembles","Not every workload shown to be riddled"],"security":"Protect training provenance logs; do not treat irreproducibility diagnostics as a reason to weaken access controls.","proofOfValue":["Measure functional churn on epsilon-close initializations","Compare GPU vs CPU deterministic baselines","Report whether finer init precision reduces churn"]},"delivery":{"work":"Pick one model family; Run controlled multi-seed / perturbation study; Update publishing acceptance checklist","dependencies":["PI lab","Research computing schedulers","Research integrity office"],"ownership":"Research computing with PI ownership of scientific claims.","skillsAdoption":"Students and staff learn to report ensembles and functional variability, not only best-seed scores.","governanceCheckpoints":["Methods must list seeds and hardware","High-stakes claims need multi-seed evidence","Periodic re-check after stack changes"],"acceptanceCriteria":["Proposed: documented churn or uncertainty-exponent style metric","Proposed: single-seed results labeled exploratory","Proposed: storage retention for replicate configs"],"risks":["Compute cost blowups","Misreading theory as a ban on DL","Ignoring simpler seed/hardware bugs"]}},"retrievedAt":"2026-11-05T22:04:45-06:00","enrichedAt":"2026-11-05T22:08:41-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Raises training and methods literacy needs for students reporting ML results; not an accessibility study.","procurementImplications":"When buying MLOps or training platforms, ask for provenance, multi-seed orchestration and reproducibility reporting features.","operatingModelImplications":"Points to an operating model where research claims from deep learning require ensemble evidence by default.","sourceVerification":{"openedUrl":"https://www.nature.com/articles/s42256-026-01336-3","referenceExcerpt":"even sharp increases in initialization precision yield only marginal gains in outcome predictability.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}