Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Precise Bayesian Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that in normalized neural networks, weight uncertainty can be compressed into one per-layer effective noise scalar with a closed-form, dimension-aware KL, yielding calibrated predictions without accuracy loss.

desk verdict Worth a serious look for its clean sigma_eff parameterization and closed-form KL, but the empirical claims are thin and the variational interpretation doesn't quite hold together as stated. read the letter →

arxiv 2506.19726 v2 pith:NPLBQPX6 submitted 2025-06-24 cs.LG cs.AIstat.ML

Carlos Stein Brito This is my paper · ORCID
classification cs.LGcs.AIstat.ML MSC 62H1168T0762F15
keywords BayesianneuralnetworksvonMises-Fisherdirectionalstatisticsvariationalinferencebatchnormalizationcalibrationeffectivenoisedimension-awareKL
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

Bayesian neural networks usually place Gaussian uncertainty on every weight, which fights the geometry of networks that use normalization. This paper claims that after normalization only the direction of a weight vector matters, so uncertainty should live on the unit sphere. It reduces each layer's uncertainty to one scalar, the effective post-normalization noise $\sigma_{\mathrm{eff}}$, and derives a closed-form dimension-aware KL penalty for it. If correct, variational BNN training becomes a small modification of standard normalized networks and improves calibration without losing accuracy.

What carries the argument

The central object is the von Mises-Fisher posterior on the unit sphere $S^{D-1}$ for weight directions, the spherical analogue of a Gaussian. It is compressed through two formulas: the activation-variance interpolant $\sigma_u^2(\kappa) \approx D/(\kappa+D)$ and the effective noise $\sigma_{\mathrm{eff}} = \sigma_u / A_D(\kappa)$, where $A_D(\kappa) = I_{D/2}(\kappa)/I_{D/2-1}(\kappa)$ is the mean resultant length. These yield one optimized scalar per normalized layer and the closed-form KL $\operatorname{KL}_{\mathrm{approx}}(\sigma_{\mathrm{eff}},D) = \frac{D-1}{2}\log\left(1 + \frac{D}{D-1}\sigma_{\mathrm{eff}}^{-2}\right)$, whose $(D-1)/2$ prefactor makes the regularization scale with layer input dimensionality.

What would settle it

Use a first-layer convolution with small input dimension, say $D=3$ for RGB images, compute the exact numerical vMF KL at intermediate concentrations, and compare it with $\operatorname{KL}_{\mathrm{approx}}$; if the gap changes the optimal $\sigma_{\mathrm{eff}}$ and ECE degrades on CIFAR-10 relative to the full pipeline, the claim that one closed-form scalar works across all normalized layers is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, a von Mises-Fisher posterior over weight directions with a uniform spherical prior is the right variational family for normalized networks. The central quantitative claim is that this posterior is equivalent, in expectation, to injecting Gaussian noise $\sigma_{\mathrm{eff}}$ after normalization, with $\sigma_{\mathrm{eff}} = \sigma_u / A_D(\kappa)$ and $\sigma_u^2 \approx D/(\kappa+D)$. The KL of the directional posterior against the uniform prior has the closed form $\operatorname{KL}_{\mathrm{approx}}(\sigma_{\mathrm{eff}},D) = \frac{D-1}{2}\log\left(1 + \frac{D}{D-1}\sigma_{\mathrm{eff}}^{-2}\right)$, which matches the true vMF KL in both tight and broad regimes. Optimizing NLL plus a weighted sum of these per-layer KL terms improves calibration, reducing ECE by roughly a factor of $5.6\times$ on CIFAR-10 with VGG16+BN, while keeping accuracy at least competitive and lowering NLL.

Load-bearing premise

The load-bearing premise is that the compact formulas $\sigma_u^2 \approx D/(\kappa+D)$ and the closed-form KL stay accurate for the whole training trajectory, including intermediate concentrations and small input dimensions; the paper validates them synthetically at $D=100$ and notes that early convolutional layers weaken the high-dimensional asymptotics.

Editorial extensions

If this is right

  • A normalized network can be trained with one extra scalar per layer and a KL warm-up, with the KL penalty scaling automatically with layer input dimension.
  • Sampling vMF weight directions is replaced by adding Gaussian noise of size $\sigma_{\mathrm{eff}}$ after batch normalization, so the training forward pass uses only standard operations.
  • At test time the noise can be turned off, leaving a deterministic network, or a few Monte Carlo passes can provide predictive uncertainty.
  • The reported CIFAR-10 VGG16+BN result predicts that this recipe lowers expected calibration error by roughly a factor of $5.6\times$ while keeping accuracy competitive and NLL lower.
  • The same geometric alignment should apply to other normalized architectures, though regression and non-BN pipelines may require minor adaptations.

Reading between the lines

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

  • A testable extension is to replace $\operatorname{KL}_{\mathrm{approx}}$ with the exact numerical vMF KL only for the lowest-dimensional layers; if calibration holds and training cost stays flat, the recipe transfers to small-$D$ layers with a hybrid objective.
  • Because $\sigma_{\mathrm{eff}}$ is just additive Gaussian noise after BN, it could be interpreted as an alternative to dropout or label smoothing; running the same VGG16 baseline with tuned Gaussian noise but no KL would isolate how much of the calibration gain comes from the KL rather than the noise injection.
  • The structured per-layer $\sigma_{\mathrm{eff}}$ profile, larger at mid-depth, hints at a cheap diagnostic for layer difficulty or a target for layer-wise learning-rate schedules, though the paper does not pursue this.
  • If the dimension-aware KL is as accurate as claimed, the $(D-1)/2$ prefactor already sets the regularization scale, so wider or deeper models could be trained without any $\beta$ search; this is a direct transfer prediction that future experiments could test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a directional Bayesian neural network method for architectures with normalization layers. It places a von Mises-Fisher (vMF) posterior on the unit sphere of weight directions, introduces approximations sigma_u^2(kappa) ≈ D/(kappa+D) and sigma_eff = sigma_u / A_D(kappa), and derives a closed-form, dimension-aware KL approximation KL_approx(sigma_eff,D) = (D-1)/2 log(1 + D/(D-1) sigma_eff^{-2}). The training objective is NLL plus a beta-weighted sum of per-layer KL terms, with one learned scalar sigma_eff per normalized layer. The forward pass injects additive Gaussian noise after batch normalization. The paper validates the approximation on synthetic data at D=100, demonstrates student-teacher parameter recovery, and reports improved calibration on CIFAR-10 with VGG16+BN, claiming about a 5.6x reduction in ECE. Limitations are acknowledged regarding small input dimensions and early-training stationarity.

Significance. The closed-form directional KL is elegant and dimension-aware, and the asymptotic-matching argument in Section 5 and Appendix F is sound; the synthetic validation at D=100 is a useful check. If the variational interpretation were established and the empirical result replicated, the method would be a lightweight, practical addition to normalized architectures. However, the manuscript does not show that Eq. (11) is an ELBO for the stochastic forward pass actually trained, and the main empirical claim rests on a single unseeded run. The significance is therefore conditional: the paper presents an attractive and potentially useful framework, but its central 'Bayesian' and 'variational' claims are not yet supported.

major comments (4)
  1. [Sections 3.4 and 5, Eq. (11)] The objective is not derived as a variational lower bound for the model actually trained. The forward pass uses fixed unit-norm weights and injects independent Gaussian noise after BN, whereas the KL term is the vMF-versus-uniform KL for a distribution over weight directions. The statement in Section 3.4 that Gaussian noise with scale sigma_eff 'matches the effect (in expectation)' of vMF sampling is a first- and second-moment match; it does not make Eq. (11) an ELBO. For the linear-Gaussian student-teacher task, moment matching suffices for expected NLL, but the CIFAR-10 experiment uses a softmax likelihood, whose expectation under the projected vMF noise depends on higher moments that the Gaussian approximation does not control. The authors should either derive an ELBO for the actual stochastic forward pass or explicitly reframe the method as a KL-regularized training scheme rather than a variational BNN.
  2. [Section 6.4 and Figure 5] The main empirical claim rests on a single unseeded run on one dataset and one architecture. There are no error bars, no reported seeds, and no comparisons with standard uncertainty baselines such as MC dropout, mean-field VI, or deep ensembles. The reported 5.6x ECE reduction is therefore not established. Please report repeated runs with variance, add at least one standard baseline, and state explicitly whether the no-KL baseline is otherwise identical in architecture, optimizer, and training schedule.
  3. [Eqs. (6)-(7) and Figure 3(A)] The interpolant D/(kappa+D) is not a close approximation to the spherical-average variance 1 - A_D(kappa)^2 in the moderate-kappa regime. Using the small-kappa expansion A_D(kappa) = kappa/D + O(kappa^3) from Appendix E, the exact variance is 1 - kappa^2/D^2 + O(kappa^4), while D/(kappa+D) = 1 - kappa/D + O(kappa^2); the linear term differs. For example, at D=100 and kappa=30 the interpolant gives about 0.77 while the leading exact term is about 0.91. Since sigma_eff and KL_approx are defined through this interpolant, the claim in Figure 3(A) that the approximation tracks ground truth needs a quantitative comparison over the kappa range actually visited during training, or the claim should be weakened.
  4. [Figure 6] The analytic vMF prediction Delta-L = kappa(1 - cos theta) requires a value of kappa, but the text does not state how kappa is determined. If kappa is fitted to the observed landscape, the agreement in Figure 6 is circular and cannot independently support the directional uncertainty model. Please report the fitting procedure and the resulting kappa value, or remove the claim that the figure validates the model.
minor comments (5)
  1. [Sections 3.3 and 4.1] The interpolant in Eq. (4) is introduced before the exact spherical-average identity in Eq. (6), which could give the impression that Eq. (4) is exact; reordering or an explicit cross-reference would clarify the status of the approximation.
  2. [Appendix A] The CIFAR-10 implementation details are incomplete: the appendix states that optimizer and schedules follow standard practice without specifying them, and no code or data availability statement is provided. This limits reproducibility.
  3. [Sections 5 and A] For convolutional layers, the notation D_l is ambiguous; it should be stated whether D_l is the fan-in, fan-in times kernel area, or the number of input channels, since the KL and sigma_eff depend directly on D_l.
  4. [Figures 3 and 4] The Monte Carlo validation plots would benefit from error bars or a statement of the number of MC samples and seeds, since the text reports that measurements 'match' without quantifying the agreement.
  5. [Abstract and Section 7] The abstract's claim that the method is 'simultaneously principled, practical, and precise' is overstated in light of the missing ELBO derivation; 'principled' should be qualified until the objective is placed on a variational footing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the vMF-based objective is built from external asymptotics and validated against numerical KL, not from its own empirical outputs.

full rationale

The paper's derivation chain is: (i) vMF posterior on the sphere with uniform prior; (ii) exact moment identities and spherical-average variance sigma_u^2 = 1 - A_D(kappa)^2; (iii) a stated approximation sigma_u^2 = D/(kappa+D) chosen to match broad- and tight-cap limits; (iv) definition sigma_eff = sigma_u/A_D(kappa); and (v) a closed-form KL_approx chosen to match the exact vMF-vs-uniform KL asymptotics. None of these steps takes its target as an input: KL_approx is validated against numerical evaluation of the exact vMF KL in Figure 3D, the variance interpolant is checked against Monte Carlo in Figure 3A, and the CIFAR-10 calibration comparison is an empirical outcome rather than a consequence of the definitions. The statement that additive Gaussian noise of scale sigma_eff 'matches the effect (in expectation)' is a moment-matching definition, not a fitted prediction. The Figure 6 loss-landscape comparison is qualitative and the paper does not state that kappa was fitted to the observed landscape, so there is no exhibited reduction of a prediction to a fit. References are external (e.g., [21] is not by the author), so no self-citation chain is load-bearing. Remaining concerns—single unseeded CIFAR-10 run, no standard BNN baselines, and the Gaussian-vs-vMF higher-moment mismatch—are correctness/empirical-support issues, not circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on three ad hoc approximations introduced specifically for this method: the activation variance interpolant, the KL surrogate, and the regime inversions. These are motivated by asymptotics and one synthetic check, but their accuracy across the full optimization range is not proven. The domain assumptions about normalization invariance and isotropic inputs are standard for normalized networks but unverified at the layer level. No new physical entities are introduced; sigma_eff is a reparameterization.

free parameters (2)
  • beta (KL weight) = warm-up from 0 to 1
    Hyperparameter controlling the strength of the KL penalty; chosen by hand, not derived. The paper uses a 20-epoch warm-up on CIFAR-10.
  • sigma_eff initial values = not specified
    Initial values for the per-layer noise scalars are not given, which affects reproducibility and can influence optimization.
assumptions (6)
  • domain assumption Normalization (BN without affine parameters) makes the pre-activation scale invariant, so the downstream function depends only on weight direction.
    Justifies restricting the posterior to the unit sphere. Stated in Section 3.1, Eq. 1, relying on BN with affine=False.
  • domain assumption Inputs after normalization are approximately isotropic with typical norm squared ||x||^2 approximately D.
    Used to derive sigma_u^2 approximately D/(kappa+D) from the spherical average identity (Section 4.1, Eq. 6). Not proven for trained networks.
  • ad hoc to paper The compact interpolant D/(kappa+D) accurately approximates the exact activation variance 1 - A_D(kappa)^2 across practical (D, kappa) ranges.
    Introduced in Section 4.2 as a special-function-free surrogate; only checked empirically at D=100 (Section 6.2), and the paper notes small input dimensions weaken high-dimensional asymptotics.
  • ad hoc to paper The closed-form KL_approx(sigma_eff,D) is an accurate surrogate for the true vMF-vs-uniform KL during optimization.
    Defined in Section 5, Eq. 10, and justified only by matching asymptotics and one Monte Carlo check (Figure 3D). No proof of accuracy in intermediate regimes.
  • standard math Bessel function identities and Laplace's method for the vMF normalizer.
    Used for exact moments and KL asymptotics in Appendices B, E, F; standard results in directional statistics.
  • standard math Spherical average identity E[x^T S x] = r^2 Tr(S)/D for x uniform on sphere.
    Used in Section 4.1 and Appendix C to derive the activation variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Precise Bayesian Neural Networks." pith.science (2026). https://pith.science/paper/NPLBQPX6

@misc{pith2026250619726,
  author       = {Pith},
  title        = {Pith review of: Precise Bayesian Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPLBQPX6}},
  note         = {Machine review of arXiv:2506.19726}
}
abstract

Despite its long history, Bayesian neural networks (BNNs) and variational training remain underused in practice: standard Gaussian posteriors misalign with network geometry, KL terms can be brittle in high dimensions, and implementations often add complexity without reliably improving uncertainty. We revisit the problem through the lens of normalization. Because normalization layers neutralize the influence of weight magnitude, we model uncertainty \emph{only in weight directions} using a von Mises-Fisher posterior on the unit sphere. High-dimensional geometry then yields a single, interpretable scalar per layer--the effective post-normalization noise $\sigma_{\mathrm{eff}}$--that (i) corresponds to simple additive Gaussian noise in the forward pass and (ii) admits a compact, dimension-aware KL in closed form. We derive accurate, closed-form approximations linking concentration $\kappa$ to activation variance and to $\sigma_{\mathrm{eff}}$ across regimes, producing a lightweight, implementation-ready variational unit that fits modern normalized architectures and improves calibration without sacrificing accuracy. This dimension awareness is critical for stable optimization in high dimensions. In short, by aligning the variational posterior with the network's intrinsic geometry, BNNs can be simultaneously principled, practical, and precise.

Figures

Figures reproduced from arXiv: 2506.19726 by the authors.

Figure 1
Figure 1. Motivational comparison. The empirical landscape is bowl-shaped with near [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Geometric view of directional weight uncertainty on the sphere. Left: vMF [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Validation of theoretical components at D = 100. (A) Activation variance: Monte Carlo vs. interpolant D/(κ + D). (B) Mean resultant length AD(κ) (Bessel ratio) vs. MC. (C) Effective post-normalization noise σeff = σu/AD(κ) alignment. (D) KL: numerical vMF￾vs-uniform vs. KLapprox(σeff, D) with asymptotes. 6 Experiments We validate the framework’s principles and core mechanics rather than pursue exhaus￾tive benchmarks… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Student-teacher recovery. Top: cosine similarity and inferred [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: CIFAR-10 (VGG16+BN). Left-to-right/top-to-bottom: accuracy, NLL, calibra [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Empirical train-loss landscapes (top) and analytic vMF predictions (bottom) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages

  1. [1]

    Dhillon, Joydeep Ghosh, and Suvrit Sra

    Anindya Banerjee, Inderjit S. Dhillon, Joydeep Ghosh, and Suvrit Sra. Clustering on the unit hypersphere using von mises-fisher distributions.Journal of Machine Learning Research, 6:1345–1382, 2005

    work page 2005

  2. [2]

    Bishop.Pattern Recognition and Machine Learning

    Christopher M. Bishop.Pattern Recognition and Machine Learning. Springer, 2006

    2006

  3. [3]

    Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M

    Tim R. Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M. Tom- czak. Hyperspherical variational auto-encoders. InAdvances in Neural Information Processing Systems 31, pp. 330–340. Curran Associates, Inc., 2018

    work page 2018

  4. [4]

    Radial bayesian neural net- works: Beyond discrete support in large-scale bayesian deep learning

    Sebastian Farquhar, Michael A Osborne, and Yarin Gal. Radial bayesian neural net- works: Beyond discrete support in large-scale bayesian deep learning. InInternational Conference on Artificial Intelligence and Statistics, pp. 1352–1362. PMLR, 2020

    work page 2020

  5. [5]

    Dropout as a bayesian approximation: Repre- senting model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Repre- senting model uncertainty in deep learning. InProceedings of The 33rd International Conference on Machine Learning, volume 48 ofPMLR, pp. 1050–1059. PMLR, 2016

    work page 2016

  6. [6]

    Hyperspherical weight uncertainty in neural net- works

    Biraja Ghoshal and Allan Tucker. Hyperspherical weight uncertainty in neural net- works. InAdvances in Intelligent Data Analysis XIX, pp. 3–14. Springer, 2021

    work page 2021

  7. [7]

    Practical variational inference for neural networks

    Alex Graves. Practical variational inference for neural networks. InAdvances in Neural Information Processing Systems 24, pp. 2348–2356. Curran Associates, Inc., 2011

    work page 2011

  8. [8]

    On calibration of modern neural networks.International Conference on Machine Learning, pp

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks.International Conference on Machine Learning, pp. 1321–1330, 2017

    work page 2017

Show all 22 references
  1. [9]

    Ioffe and C

    S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. InProceedings of the International Conference on Machine Learning, 2015

    2015

  2. [10]

    What are bayesian neural network posteriors really like?International Conference on Machine Learning, pp

    Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Wilson. What are bayesian neural network posteriors really like?International Conference on Machine Learning, pp. 4629–4640, 2021

    2021

  3. [11]

    Jordan, Zoubin Ghahramani, Tommi S

    Michael I. Jordan, Zoubin Ghahramani, Tommi S. Jaakkola, and Lawrence K. Saul. An introduction to variational methods for graphical models.Machine Learning, 37(2): 183–233, 1999

    1999

  4. [12]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. InProceedings of the 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014

    2014

  5. [13]

    Kingma, Tim Salimans, and Max Welling

    Diederik P. Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. InAdvances in Neural Information Processing Systems 28, pp. 2575–2583. Curran Associates, Inc., 2015

    2015

  6. [14]

    Mardia and Peter E

    Kanti V. Mardia and Peter E. Jupp.Directional Statistics. John Wiley & Sons, 2000

    2000

  7. [15]

    G-SGD: Optimizing ReLU neural networks in its positively scale-invariant space

    Qi Meng, Shuxin Zheng, Huishuai Zhang, Wei Chen, Zhi-Ming Ma, and Tie-Yan Liu. G-SGD: Optimizing ReLU neural networks in its positively scale-invariant space. In International Conference on Learning Representations (ICLR), 2019

    2019

  8. [16]

    Variational dropout spar- sifies deep neural networks

    Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout spar- sifies deep neural networks. InProceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 ofProceedings of Machine Learning...

    2017

  9. [17]

    Calibrating deep neural networks using focal loss.Advances in Neural Information Processing Systems, 33:15744–15755, 2020

    Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss.Advances in Neural Information Processing Systems, 33:15744–15755, 2020

    2020

  10. [18]

    Norm-based capacity con- trol in neural networks

    Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity con- trol in neural networks. InProceedings of The 28th Conference on Learning Theory, volume 40 ofProceedings of Machine Learning Research, pp. 1376–1401. PMLR, 2015

    2015

  11. [19]

    Radial and directional pos- teriors for bayesian deep learning.Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5298–5305, 2020

    Changyong Oh, Kamil Adamczewski, and Mijung Park. Radial and directional pos- teriors for bayesian deep learning.Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5298–5305, 2020

    2020

  12. [20]

    Salimans and D

    T. Salimans and D. P. Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. InAdvances in Neural Information Processing Systems, 2016

    2016

  13. [21]

    Directional statistics in machine learning: A brief review.arXiv preprint arXiv:1605.00316, 2016

    Sanjoy Sra. Directional statistics in machine learning: A brief review.arXiv preprint arXiv:1605.00316, 2016

    2016 arXiv

  14. [22]

    Tensor programs v: Tuning large neural networks viaµ-parametrization

    Greg Yang. Tensor programs v: Tuning large neural networks viaµ-parametrization. arXiv preprint arXiv:2203.03466, 2022. A Implementation details Parameterization and forward.Each layer maintains a single positive scalarσ eff (we optimizeρwithσ eff = softplus(ρ) or exp(ρ)). Giv...

    2022 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.