Pith. sign in

REVIEW 5 major objections 6 minor 32 references

Neural Feature Governance: Extending Atom Prevalence

T0 review · 5 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Neural Atom Prevalence prunes feedforward networks to as little as 8% of their original nodes while keeping accuracy near dense baselines and model uncertainty near 3–4% of predictive variance.

desk verdict A promising but overclaimed hybrid pipeline for neuron-level Bayesian compression; the theoretical and calibration claims need to be reined in before it's citable. read the letter →

arxiv 2607.21671 v1 pith:XSRUIIBI submitted 2026-07-23 cs.LG

Idris Karel Seunda Ekwe , Patrick Tenga Shako , Ernest Parfait Fokoué This is my paper
classification cs.LG
keywords BayesianneuralnetworkatomprevalencespikeandslabuncertaintyquantificationPoisson-Binomialselectioncompressionvariationalinferencelotterytickethypothesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to extend the atom-prevalence selection principle — first estimate the most likely model size, then keep the most probable atoms — from linear regression to hidden units of feedforward neural networks. The proposed Neural Atom Prevalence (NAP) pipeline finds a lottery-ticket subnetwork by iterative magnitude pruning, trains a spike-and-slab variational network with a soft Gumbel-Softmax relaxation, computes a posterior inclusion probability for every neuron, and uses the mode of the Poisson-Binomial layer-size distribution to decide how many neurons to keep. The paper's central claim is that this hybrid procedure yields networks that are simultaneously much smaller — 56–92% of nodes removed across simulated regression, two real-world regression benchmarks, and MNIST — and nearly as accurate as dense baselines, while remaining probabilistically well-calibrated, with epistemic uncertainty at roughly 3–4% of total predictive variance and 93.4% observed coverage of nominal 95% intervals. A sympathetic reader would care because the method offers one coherent selection rule for compression, interpretability, and uncertainty quantification at once, rather than treating pruning and calibration as separate engineering steps.

What carries the argument

The load-bearing mechanism is the neural atom selection rule: rank neurons in each layer by their variational inclusion probabilities and keep the top K_opt^l, where K_opt^l is the mode of the Poisson-Binomial posterior for that layer's size. This replaces an arbitrary pruning threshold with a probability-theoretic answer to how many and which neurons to keep. The supporting machinery includes the lottery-ticket subnetwork found by iterative magnitude pruning, which sets the prior inclusion probabilities and the starting topology; the soft Gumbel-Softmax relaxation used in both forward and backward passes to avoid premature pruning; and the epoch-0 rewinding of the final sparse network befor

What would settle it

Rerun the simulated-data experiment with the Poisson-Binomial atom-prevalence selection replaced by random pruning at the same K_opt layer sizes, keeping all other NAP phases fixed. If random pruning reproduces the reported RMSE (~1.20), 93.4% interval coverage, and ~3% epistemic fraction, then the atom-prevalence ranking is not what drives the results; if random pruning clearly fails on those metrics, the selection rule is doing real work.

Watch

Extended reading notes

Core claim

The core transposition is to call a neuron, defined as the activation unit h_j^l = ψ_l(g_j^l), a neural atom and to apply the atom-prevalence logic to it. A spike-and-slab prior groups all incoming weights and bias of each neuron, with a Dirac spike for exclusion and a Gaussian slab for inclusion; variational inference over the inclusion indicators yields a posterior inclusion probability for every neuron. The layer size is then treated as a Poisson-Binomial random variable, and the optimal layer size is its mode, K_opt^l = arg max_k P(K_V^l = k), computed by dynamic programming; the final network keeps the K_opt^l neurons with the highest inclusion probabilities in each layer, then fine-tun

Load-bearing premise

The load-bearing premise is that a mathematical guarantee proved for a simpler earlier method still holds after the NAP pipeline changes the training procedure — soft Gumbel-Softmax in both passes, lottery-ticket-based prior calibration, Poisson-Binomial mode selection, and epoch-0 rewinding — and the paper does not re-prove that guarantee for these modifications; if any change breaks the earlier proof's conditions, the claim that NAP is theoretically grounded rests on a heur

Editorial extensions

If this is right

  • If NAP works as claimed, feedforward networks can lose 56–92% of their nodes while staying within one to two accuracy points, or tenths of RMSE, of dense baselines.
  • Regression uncertainty from the pruned network is nearly calibrated: roughly 3–4% epistemic variance, 93.4% observed coverage of nominal 95% intervals, and regression calibration errors below 0.05.
  • The Poisson-Binomial mode gives a data-dependent, probability-based layer size, so a practitioner no longer needs an arbitrary pruning fraction or magnitude threshold.
  • The winning-ticket initialization can improve accuracy over both dense variational training and standard spike-and-slab pruning on some datasets, most clearly on the Concrete regression benchmark, where NAP reports RMSE 6.32 versus 7.92 and 7.34 for baselines.
  • The calibration story does not extend to classification as cleanly: MNIST shows overconfidence in low-to-mid confidence bins, so classification deployment would need post-hoc calibration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's consistency guarantee is inherited from an earlier layer-adaptive spike-and-slab method, not re-proved for the exact NAP pipeline; a natural test is to remove the lottery-ticket phase and check whether accuracy, coverage, and sparsity stay fixed. If they do, magnitude pruning rather than atom-prevalence ranking is carrying the compression.
  • The same rank-by-inclusion-probability and keep-the-Poisson-Binomial-mode recipe should transfer to convolutional or transformer layers by treating output channels as atoms; nothing in the selection logic depends on layers being fully connected, although the paper only tests feedforward networks.
  • Because pruning drives epistemic uncertainty down to roughly 3–4% in-distribution, the model will likely be overconfident on shifted inputs; the paper identifies out-of-distribution behavior as an open weakness, and entropy-on-shifted-input is the direct next experiment.
  • Choosing the Poisson-Binomial mode as the optimal layer size is carried by analogy from linear atom prevalence rather than proven optimal for layered nonlinear features; comparing it with a median or Bayesian-model-averaged layer size on small samples would be a cheap check.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Neural Atom Prevalence (NAP), a four-phase pipeline for feedforward BNNs: Bayesian Lottery Ticket identification via iterative magnitude pruning, soft Spike-and-Slab Independent Gaussian (SS-IG) variational training, Poisson-Binomial layer-size selection, and Bayesian fine-tuning from the epoch-0 initialization with π=0.99. NAP treats activation units as 'neural atoms' and claims to extend Fokoué's atom-prevalence principle from linear predictors to neurons. The core theoretical claim is that the variational posterior Q* contracts at the minimax-optimal rate by inheriting results from Jantre et al. (2023). Empirical validation is reported on simulated nonlinear regression, Concrete, YearPredictionMSD, and MNIST, with claims of 56–92% node removal, accuracy close to dense baselines, 93.4% interval coverage versus a 95% target, and a 3–4% epistemic fraction of total predictive variance.

Significance. If the transfer of atom prevalence to neural atoms were rigorously established, NAP would be a useful hybrid of lottery-ticket pruning and Bayesian node selection, with an explicit neuron-level selection rule and a tractable layer-size posterior. The manuscript has genuine strengths: a clearly specified pipeline, detailed hyperparameter tables, multiple benchmarks, fixed seeds, and a candid list of limitations in the conclusion. These are positive features. However, the central 'theoretically grounded' claim is currently an inheritance assertion, not a derivation, and the empirical calibration story is contradicted by the manuscript's own MNIST results. The paper would be valuable as an empirical pruning study, but in its present form the theoretical framing overstates what is established.

major comments (5)
  1. [§3.3.2, Eq. (12); §3.5] The consistency claim is made for Q*, the variational posterior of the soft SS-IG phase, but the final NAP model is a different estimator: after PB selection, non-selected neurons are discarded and the surviving network is rewound and fine-tuned with π=0.99, forcing all remaining nodes active. No theorem in the paper or in the cited literature is shown to cover this composed estimator. The listed conditions from Jantre et al. do not address the modified training procedure (soft Gumbel-Softmax in both passes, BLT-based π_l, hard selection, rewinding). The abstract's 'theoretically grounded' claim therefore rests on an unverified inheritance. Either prove contraction for the final pipeline or reposition the method as heuristic.
  2. [§3.4.3–§3.4.4; Remark 1] The rule K_opt^l = argmax_k P(K_V^l = k), followed by selecting the top K_opt^l inclusion probabilities, is asserted to be optimal by analogy with Fokoué's linear atom-prevalence result. No proof is given for layered nonlinear features, where inter-layer dependencies and nonlinear composition matter. Moreover, Remark 1 requires at least one active neuron per layer, but the Poisson-Binomial mode can be k=0 when inclusion probabilities are small, and the selection rule imposes no constraint preventing an invalid empty layer. The selection rule needs either an explicit per-layer non-collapse constraint or a proof that the unconstrained mode is always ≥1 under the proposed π_l calibration.
  3. [§3.4.1, Eq. (20)] The calibration of π_l is not derived. Substituting winning-ticket statistics S_wt and B_wt into Jantre et al.'s Corollary 4.5 is an assertion; no argument shows that the corollary's hypotheses are preserved after IMP pruning. The stated constraint '10^-50 ≤ π_l ≤ S_wt' is dimensionally invalid as written, because S_wt is defined as an l0 count of L1 norms, not a probability. Since π_l controls the whole prior and the PB layer-size distribution, this is load-bearing for the selection procedure. A precise, valid calibration rule is needed.
  4. [§3.1, §3.3.2] The consistency theorem invoked in §3.3.2 requires 1-Lipschitz activation functions, but all experiments use Swish/SiLU. The derivative of f(x)=x·σ(x) reaches ≈1.0998, so Swish is not 1-Lipschitz. The paper's statement that Swish satisfies the condition is incorrect. Thus even the inherited theorem does not apply to the actual experimental setup. Either use a genuinely 1-Lipschitz activation in the SS-IG phase or extend the theorem to the activation used.
  5. [§4.3.1, §4.3.3, §4.4.2 vs. Abstract] The abstract's unqualified 'well-calibrated probabilistically' and the claim that epistemic uncertainty is 3–4% of total predictive variance 'across all experiments' are contradicted by the manuscript's own numbers. MNIST reports ECE=0.1005 and MCE=0.3603 with systematic overconfidence in the 0.2–0.6 confidence bins. The 93.4% coverage and calibration errors 0.0273–0.0310 are regression-only. Furthermore, the aleatoric/epistemic decomposition is never specified for the classification model: §3.1 fixes the Gaussian likelihood with σ_e²=1, yet the results report estimated aleatoric means such as 1.2456 and 8.213. The model for the observation noise used in the decomposition is not given, so the 3–4% claim is unverifiable.
minor comments (6)
  1. [Figures] The manuscript text contains only figure captions; the actual panels A–F are not visible in the submitted version. Without the plots, the calibration diagrams and uncertainty decompositions cannot be checked.
  2. [§2] There is an inconsistent citation name: the text says 'Frankle and Carlin' but the reference list and the standard literature use 'Frankle and Carbin.'
  3. [§4.2.2] The text says MNIST pixel values are divided by 126. The usual normalization is 255; if this is intentional, please justify it, otherwise correct the typo.
  4. [Throughout] The manuscript repeatedly refers to 'this thesis' and uses thesis-style language. For a journal article, use 'this paper' throughout and adjust the framing.
  5. [§5] The conclusion reports 'MACE = 0.0131', but MACE is not defined in §4.2.6 and does not appear in the experimental sections. Please define it or remove it.
  6. [Abstract / §4.5] The abstract calls the sparsity 'state-of-the-art,' but the comparison table includes only SS-IG and VBNN, not standard magnitude-pruning or other neural-network compression baselines. The 'state-of-the-art' wording is unsupported by the presented comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: NAP's sparsity, accuracy, and coverage figures are post-hoc outputs of its own trained posterior, not fitted targets; the consistency gap in Section 3.3.2 is an unverified theorem transfer, not a definitional circle.

full rationale

Walking the derivation chain, I find no step in which a reported prediction is equivalent to an input by construction. The inclusion probabilities pi_tilde_j^l, the Poisson-Binomial layer-size posterior P(K_V^l=k), the mode K_opt^l, and the top-inclusion selection rule are all computed from the model's own trained variational posterior; none is fitted to test accuracy, RMSE, or interval coverage, so the claimed sparsity and calibration numbers are genuine post-hoc evaluations rather than forced identities. The selection rule is transparently inherited from Fokoue (2008), a coauthor's prior work, and the variational consistency result is borrowed from Jantre et al. (2023); these are external methods, and the paper's empirical comparison supplies independent content. The most serious concern is not circularity but an unproven transfer: Section 3.3.2 asserts that Q*'s contraction 'directly justify[ies] the consistency claim for the NAP sparse architecture whose topology varies across the IMP, SS-IG, and fine-tune phases,' yet Algorithm steps 9-10 produce a different estimator - PB-pruned, rewound to epoch-0, and fine-tuned with pi=0.99 forcing all surviving nodes active - and no cited theorem covers that composed object. Likewise, the abstract's unqualified 'well-calibrated' statement conflicts with the paper's own reported MNIST ECE=0.1005 and MCE=0.3603. These are rigor/correctness gaps, not definitional circularity, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 7 assumptions · 1 invented entities

The central method depends on a chain of borrowed theoretical results (Fokoué's prevalence rule, Jantre et al.'s SS-IG consistency) plus several hand-chosen hyperparameters: π_l calibration, τ=0.5, IMP rounds/fraction, KL annealing, and rewinding with π=0.99. There are no new physical entities; 'neural atom' is a relabeling of neuron activations. The most load-bearing unproven step is the transfer of Jantre et al.'s consistency theorem to the modified NAP pipeline.

free parameters (6)
  • Layer-wise prior inclusion probabilities π_l = Computed from winning-ticket statistics via Corollary 4.5 of Jantre et al.; C_l chosen so 1e-50 ≤ π_l ≤ S_wt
    The prior inclusion probability for every node in a layer is calibrated using the winning ticket's surviving node count and L1 norms, with the constant C_l chosen by hand to keep π_l in an arbitrary range. This prior directly shapes which nodes are selected.
  • Gumbel-Softmax temperature τ = 0.5
    Chosen by hand as 'the most suitable value' with no sensitivity analysis; it controls the soft/discrete tradeoff during training and affects the posterior inclusion estimates.
  • IMP rounds and prune fraction = 5 rounds, 0.35 per round
    Chosen by hand; determines the winning-ticket topology, which in turn sets the prior calibration and the starting architecture for SS-IG training.
  • KL annealing schedule = Linear 0→1 over first 400 epochs
    Arbitrary schedule inherited from the SS-IG setup; no sensitivity analysis is given, yet it affects the quality of the variational posterior and the resulting inclusion probabilities.
  • Rewinding prior inclusion π=0.99 for fine-tuning = 0.99
    Design choice to render all retained nodes active at initialization of the fine-tuning phase; justified only by analogy to the Lottery Ticket Hypothesis, with no ablation.
  • MC samples at test time = 10
    Only 10 Monte Carlo samples are used for predictive uncertainty; no convergence check is reported, so the reported coverage and calibration numbers carry Monte Carlo noise.
assumptions (7)
  • standard math The Poisson-Binomial dynamic programming recursion of Hong (2013) computes the exact layer-size distribution.
    Used in §3.4.2 to compute P(K_V^l = k) with O(K^2_max) complexity. This is a standard algorithm, not in dispute.
  • domain assumption The modeling assumptions σ_e²=1, X uniform on [0,1], and 1-Lipschitz activation functions hold.
    Invoked in §3.1 following Jantre et al. (2023). If these assumptions are violated, the cited consistency theorem's conditions may not hold for the NAP pipeline.
  • domain assumption At least one neuron must remain active in each hidden layer for the network to define a valid function.
    Remark 1 in §3.1.2. This is a sensible structural constraint, but it also changes the prior model count from 2^{sum K_l} to ∏(2^{K_l}-1), which is handled only informally.
  • domain assumption The mean-field variational family with Gumbel-Softmax approximation is adequate for posterior inference over node inclusion.
    Invoked throughout §3.3. The paper itself notes in §5 that mean-field systematically underestimates posterior variance, directly weakening the calibration claims.
  • ad hoc to paper Winning-ticket statistics (S_wt, B_wt) can be substituted into Jantre et al.'s Corollary 4.5 to obtain a valid prior inclusion probability π_l.
    Section 3.4.1 makes this substitution 'to have a more realistic π_l', but no proof is offered that the theorem still holds when the dense-network bounds are replaced by ticket-based bounds.
  • ad hoc to paper The Poisson-Binomial mode K_opt = argmax_k P(K_V^l=k) is the statistically optimal layer size.
    Section 3.4.3 asserts this by analogy to Fokoué's linear-model atom prevalence. No derivation is given for layered nonlinear neural features.
  • ad hoc to paper Rewinding the final BNN to the epoch-0 SS-IG initialization with π=0.99 is beneficial for the fine-tuning phase.
    Section 4.2.7 adopts the LTH rewinding principle without an ablation study; this choice is load-bearing for the final fine-tuning stability claims.
invented entities (1)
  • Neural atom (activation unit h_l^j)
    purpose: Reifies each neuron as an atomic selection unit, reducing model selection from weight space to node space.
    Definition 1 in §3.1.2. This is a conceptual relabeling of standard neuron activations; it has no falsifiable handle outside the method's own performance, so it is an invented formal entity rather than an empirically established one.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Feature Governance: Extending Atom Prevalence." pith.science (2026). https://pith.science/paper/XSRUIIBI

@misc{pith2026260721671,
  author       = {Pith},
  title        = {Pith review of: Neural Feature Governance: Extending Atom Prevalence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XSRUIIBI}},
  note         = {Machine review of arXiv:2607.21671}
}
read the original abstract

Neural network compression and interpretability remain open challenges in modern deep learn- ing, where billion-parameter architectures deliver impressive accuracy at the cost of trans- parency, computational efficiency, and reliable uncertainty quantification. This paper introduces Neural Atom Prevalence (NAP), a principled Bayesian framework for structured node-level model selection in feedforward neural networks. NAP introduces the neural atom (activation unit) and functions as a hybrid method operating through a four-phase pipeline: Bayesian Lottery Ticket (BLT) identification via Iterative Magnitude Pruning (IMP), soft variational training of the Spike and Slab Independent Gaussian (SS-IG) model, Poisson-Binomial (PB) optimal layer-size selection, and Bayesian fine-tuning to produce a sparse, stable, interpretable, and accurate model. Extensive empirical validation across simulated nonlinear regression, two UCI benchmark datasets (Concrete, YearPredictionMSD), and the MNIST image classification task demonstrates that NAP achieves state-of-the-art structural sparsity, reducing active nodes to as few as 8% of the original dense architecture on MNIST, while well-calibrated probabilisti- cally: the aleatoric-epistemic uncertainty decomposition reveals that model ignorance accounts for only 3 to 4% of total predictive variance across all experiments, and regression reliability diagrams confirm a near-nominal predictive interval coverage (93.4% observed against a 95% target). These results establish NAP as a reliable, theoretically grounded, and computation- ally tractable solution to the simultaneous pursuit of sparsity, accuracy, interpretability, and uncertainty quantification in Bayesian neural networks.

Figures

Figures reproduced from arXiv: 2607.21671 by the authors.

Figure 1
Figure 1. Results of experiment on simulated data:(A) RMSE vs epoch, NAP vs baseline. (B) Node [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Results of experiment on YearPredictionMSD dataset: (A) RMSE vs epoch, SS-IG phase. [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Results of experiment on Concrete Dataset:(A) RMSE vs epoch, SS-IG and fine-tune. [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Results of experiment on MNIST dataset:(A) Accuracy vs epoch, SS-IG phase. (B) Accu [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 1 canonical work pages

  1. [1]

    Computational Statistics & Data Analysis , volume=

    On computing the distribution function for the Poisson binomial distribution , author=. Computational Statistics & Data Analysis , volume=. 2013 , publisher=

    2013

  2. [2]

    2023 , issn =

    Layer adaptive node selection in Bayesian neural networks: Statistical guarantees and implementation details , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.neunet.2023.08.029 , url =

    work page doi:10.1016/j.neunet.2023.08.029 2023

  3. [3]

    Estimation of Atom Prevalence for Optimal Prediction , journal =

    Fokou. Estimation of Atom Prevalence for Optimal Prediction , journal =. 2008 , publisher =

    2008

  4. [4]

    International Conference on Learning Representations (ICLR) , year=

    The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks , author=. International Conference on Learning Representations (ICLR) , year=

  5. [5]

    arXiv preprint arXiv:2602.18825 , year=

    Bayesian Lottery Ticket Hypothesis , author=. arXiv preprint arXiv:2602.18825 , year=

    arXiv

  6. [6]

    , title =

    Han, Song and Pool, Jeff and Tran, John and Dally, William J. , title =. Advances in Neural Information Processing Systems (NIPS) , volume =. 2015 , publisher =

    2015

  7. [7]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Spike-and-Slab Shrinkage Priors for Structurally Sparse Bayesian Neural Networks , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2025 , publisher=

    2025

  8. [8]

    International Conference on Machine Learning , pages=

    Winning the lottery ahead of time: Efficient early network pruning , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

Show all 32 references
  1. [9]

    Neural Networks , volume=

    Multilayer feedforward networks are universal approximators , author=. Neural Networks , volume=. 1989 , publisher=

    1989

  2. [10]

    International Conference on Machine Learning , pages=

    Weight uncertainty in neural networks , author=. International Conference on Machine Learning , pages=. 2015 , organization=

    2015

  3. [11]

    Journal of the American Statistical Association , volume=

    Bayesian variable selection in linear regression , author=. Journal of the American Statistical Association , volume=. 1988 , publisher=

    1988

  4. [12]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    Regression shrinkage and selection via the lasso , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1996 , publisher=

    1996

  5. [13]

    The Journal of Machine Learning Research , volume=

    On model selection consistency of Lasso , author=. The Journal of Machine Learning Research , volume=. 2006 , publisher=

    2006

  6. [14]

    Journal of the American Statistical Association , volume=

    The bayesian lasso , author=. Journal of the American Statistical Association , volume=. 2008 , publisher=

    2008

  7. [15]

    Journal of the American Statistical Association , volume=

    Variable selection via Gibbs sampling , author=. Journal of the American Statistical Association , volume=. 1993 , publisher=

    1993

  8. [16]

    Machine Learning , volume=

    Random forests , author=. Machine Learning , volume=. 2001 , publisher=

    2001

  9. [17]

    BMC Bioinformatics , volume=

    Bias in random forest variable importance measures: Illustrations, sources and a solution , author=. BMC Bioinformatics , volume=. 2007 , publisher=

    2007

  10. [18]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Stability selection , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2010 , publisher=

    2010

  11. [19]

    Advances in Neural Information Processing Systems (NIPS) , volume=

    Optimal brain damage , author=. Advances in Neural Information Processing Systems (NIPS) , volume=

  12. [20]

    Journal of Machine Learning Research , volume=

    Sparsity in Deep Learning: A Comprehensive Survey across Architectures and Applications , author=. Journal of Machine Learning Research , volume=

  13. [21]

    1996 , publisher=

    Bayesian Learning for Neural Networks , author=. 1996 , publisher=

    1996

  14. [22]

    International Conference on Machine Learning (ICML) , pages=

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning , author=. International Conference on Machine Learning (ICML) , pages=. 2016 , organization=

    2016

  15. [23]

    International Conference on Machine Learning (ICML) , pages=

    Variational dropout sparsifies deep neural networks , author=. International Conference on Machine Learning (ICML) , pages=. 2017 , organization=

    2017

  16. [24]

    2024 , howpublished =

    What are Bayesian Neural Networks? , author =. 2024 , howpublished =

    2024

  17. [25]

    URL http://archive

    UCI machine learning repository, 2017 , author=. URL http://archive. ics. uci. edu/ml , volume=

    2017

  18. [26]

    Advances in neural information processing systems , volume=

    What uncertainties do we need in bayesian deep learning for computer vision? , author=. Advances in neural information processing systems , volume=

  19. [27]

    International conference on machine learning , pages=

    On calibration of modern neural networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  20. [28]

    5th International Conference on Learning Representations (ICLR) , year =

    Categorical Reparameterization with Gumbel-Softmax , author =. 5th International Conference on Learning Representations (ICLR) , year =

  21. [29]

    5th International Conference on Learning Representations (ICLR) , year =

    The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables , author =. 5th International Conference on Learning Representations (ICLR) , year =

  22. [30]

    Stochastic Environmental Research and Risk Assessment , volume=

    Statistics for sample splitting for the calibration and validation of hydrological models , author=. Stochastic Environmental Research and Risk Assessment , volume=. 2018 , publisher=

    2018

  23. [31]

    arXiv preprint arXiv:1212.2512 , year=

    A generalized mean field algorithm for variational inference in exponential families , author=. arXiv preprint arXiv:1212.2512 , year=

    arXiv

  24. [32]

    The Annals of Statistics , year=

    Optimal predictive model selection , author=. The Annals of Statistics , year=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.