arxiv: 2605.14146 · v1 · pith:VZYYFHU7new · submitted 2026-05-13 · 💻 cs.LG

bde: A Python Package for Bayesian Deep Ensembles via MILE

Vyron Arvanitis , Angelos Aslanidis , Emanuel Sommer , David R\"ugamer This is my paper

Pith reviewed 2026-05-15 04:50 UTC · model grok-4.3

classification 💻 cs.LG

keywords Bayesian deep ensemblesMicrocanonical Langevin Ensemblestabular datauncertainty quantificationPython packageJAXscikit-learnMCMC sampling

0 comments

The pith

The bde Python package supplies scikit-learn compatible estimators for Bayesian deep ensembles on tabular data via efficient JAX implementation of Microcanonical Langevin Ensembles sampling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents bde as a user-friendly Python package focused on Bayesian deep ensembles for tabular data. It implements Microcanonical Langevin Ensembles, or MILE, in JAX to enable fast training, Markov Chain Monte Carlo sampling, and uncertainty quantification for both regression and classification. The estimators follow scikit-learn conventions so they slot into standard workflows. A sympathetic reader would care because the package lowers the computational and integration cost of obtaining posterior samples and calibrated uncertainty from deep models on structured data.

Core claim

The paper claims that an efficient JAX implementation of Microcanonical Langevin Ensembles supplies scikit-learn compatible estimators that deliver fast training, efficient Markov Chain Monte Carlo sampling, and uncertainty quantification for Bayesian deep ensembles in regression and classification tasks on tabular data.

What carries the argument

Microcanonical Langevin Ensembles (MILE), a sampling-based inference method whose JAX implementation performs efficient Markov Chain Monte Carlo sampling for Bayesian deep ensembles.

If this is right

Fast training and MCMC sampling become practical for Bayesian deep learning on tabular datasets.
Uncertainty quantification is supplied for both regression and classification problems.
Scikit-learn compatibility allows direct use inside existing machine learning pipelines.
Efficient sampling supports posterior inference without prohibitive computational overhead.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The package could support uncertainty-aware decisions in domains such as finance or medicine that depend on tabular inputs.
Users might combine bde estimators with existing scikit-learn model selection tools to tune ensemble size or sampling parameters.
Future work could measure wall-clock speedups against standard Hamiltonian Monte Carlo on the same tabular tasks.

Load-bearing premise

MILE sampling produces effective Bayesian posterior approximations for deep neural networks on tabular data.

What would settle it

A benchmark comparison on standard tabular datasets in which bde ensembles show worse calibration or higher predictive error than ordinary deep ensembles or alternative Bayesian methods.

read the original abstract

bde is a user-friendly Python package for Bayesian Deep Ensembles with a particular focus on tabular data. Built on an efficient JAX implementation of the sampling-based inference method Microcanonical Langevin Ensembles (MILE), it provides scikit-learn compatible estimators for fast training, efficient Markov Chain Monte Carlo sampling, and uncertainty quantification in both regression and classification tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a software announcement for a scikit-learn wrapper around an existing JAX implementation of MILE sampling, aimed at making Bayesian deep ensembles easier for tabular data.

read the letter

The main thing here is a new Python package called bde that packages Microcanonical Langevin Ensembles (MILE) into scikit-learn compatible estimators. It targets regression and classification on tabular data, with claims of fast training, efficient MCMC sampling, and built-in uncertainty quantification. The work is incremental rather than foundational: it takes a prior sampling method and adds a convenient API layer on top of JAX. That wrapper is what is actually new, and it seems useful for people who already like scikit-learn workflows but want Bayesian-style uncertainty without starting from scratch. The focus on tabular data is a reasonable choice since most deep ensemble papers lean toward images or text. The implementation choices look practical on paper, with emphasis on efficiency through JAX. Soft spots are straightforward. The abstract and description give no benchmarks, no runtime numbers, no comparison to alternatives like standard deep ensembles or other Bayesian methods, and no verification that the sampling actually mixes well on real tabular problems. Everything rests on the assumption that MILE works as advertised, which the package does not re-test or extend. There is also no discussion of limitations, such as how the method scales with network size or handles missing data common in tabular settings. This paper is for practitioners who need a drop-in tool for uncertainty on tables rather than for readers hunting new theory or large-scale empirical results. It is a solid software contribution if the full code and examples hold up, but it does not move the underlying science forward. I would send it to peer review for the software track, with the expectation that reviewers will ask for concrete performance numbers and usage guidance.

Referee Report

0 major / 2 minor

Summary. The manuscript introduces the bde Python package, which supplies scikit-learn-compatible estimators for Bayesian deep ensembles on tabular data. It wraps an efficient JAX implementation of Microcanonical Langevin Ensembles (MILE) to enable fast training, MCMC sampling, and uncertainty quantification for both regression and classification tasks.

Significance. If the thin wrapper layer around the existing MILE implementation is correct, the package lowers the barrier to Bayesian inference for deep models on tabular data by providing a familiar scikit-learn API. The availability of the actual Python package itself is a concrete strength for reproducibility and immediate usability.

minor comments (2)

The manuscript would benefit from a short table or paragraph explicitly listing the supported estimator classes (e.g., BDERegressor, BDEClassifier) and their key hyperparameters so readers can assess the API surface without installing the package.
Consider adding a brief note on how the JAX-based sampler is exposed through the scikit-learn fit/predict interface, including any constraints on model architecture or data types that users must respect.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive review, recognition of the package's practical value for tabular Bayesian inference, and recommendation to accept the manuscript.

Circularity Check

0 steps flagged

No significant circularity; software wrapper around prior MILE method

full rationale

The paper is a package announcement for bde, which provides scikit-learn compatible wrappers around an existing JAX implementation of Microcanonical Langevin Ensembles (MILE). The abstract and description contain no new theoretical derivations, equations, fitted parameters renamed as predictions, or self-citation chains that reduce the central claim to its own inputs. The load-bearing content is documentation of API choices and implementation around a pre-existing sampling method, with no internal reduction by construction or uniqueness theorem imported from the authors' prior work. This is a standard, self-contained software paper with no circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The package depends on the pre-existing MILE inference method and standard libraries (JAX, scikit-learn) without introducing new free parameters, axioms, or invented entities in the abstract description.

pith-pipeline@v0.9.0 · 5354 in / 1119 out tokens · 38064 ms · 2026-05-15T04:50:48.482007+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

The Thirteenth International Conference on Learning Representations , year=

Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks , author=. The Thirteenth International Conference on Learning Representations , year=

work page
[2]

Proceedings of the 41st International Conference on Machine Learning , year=

Connecting the Dots: Is Mode-Connectedness the Key to Feasible Sample-Based Inference in Bayesian Neural Networks? , author=. Proceedings of the 41st International Conference on Machine Learning , year=

work page
[3]

Proceedings of the 41st International Conference on Machine Learning , year =

Papamarkou, Theodore and Skoularidou, Maria and Palla, Konstantina and Aitchison, Laurence and Arbel, Julyan and Dunson, David and Filippone, Maurizio and Fortuin, Vincent and Hennig, Philipp and Hern\'. Proceedings of the 41st International Conference on Machine Learning , year =

work page
[4]

2024, BlackJAX: Composable Bayesian inference in JAX, arXiv:2402.10797

Alberto Cabezas and Adrien Corenflos and Junpeng Lao and Rémi Louf , year=. 2402.10797 , archivePrefix=

work page arXiv
[5]

James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and George Necula and Adam Paszke and Jake Vander

work page
[6]

The Journal of Machine Learning Research , volume=

Microcanonical Hamiltonian Monte Carlo , author=. The Journal of Machine Learning Research , volume=. 2023 , publisher=

work page 2023
[7]

Symposium on Advances in Approximate Bayesian Inference , pages=

Fluctuation without dissipation: Microcanonical Langevin Monte Carlo , author=. Symposium on Advances in Approximate Bayesian Inference , pages=. 2024 , organization=

work page 2024
[8]

International Conference on Learning Representations , year=

Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

work page
[9]

and Varoquaux, G

Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V. and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P. and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E. , journal=. Scikit-learn: Machine Learning in

work page
[10]

UCI Machine Learning Repository

Dua, Dheeru and Graff, Casey. UCI Machine Learning Repository. 2017

work page 2017
[11]

2013 , howpublished =

Fanaee-T,Hadi , title =. 2013 , howpublished =

work page 2013