Variational meta-learning inference for low dimensional neural system identification

Dario Piga; Marco Forgione; Matteo Rufolo

REVIEW 4 major objections 4 minor 19 references

Variational meta-learning gives neural system identification calibrated uncertainty in low-data regimes

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 13:49 UTC pith:U2MN45IX

load-bearing objection Worth reading: a clean, self-aware probabilistic extension of the authors' deterministic manifold meta-learning; the Laplace unimodality caveat is real and central, so it is a conditional accept rather than a clear one. the 4 major comments →

arxiv 2607.18965 v1 pith:U2MN45IX submitted 2026-07-21 cs.LG cs.AIcs.SYeess.SY

Variational meta-learning inference for low dimensional neural system identification

Matteo Rufolo , Dario Piga , Marco Forgione This is my paper

classification cs.LG cs.AIcs.SYeess.SY

keywords variational inferencemeta-learningsystem identificationuncertainty quantificationLaplace approximationlow-dimensional manifoldBouc-Wen benchmark

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a fully probabilistic reformulation of manifold meta-learning can quantify epistemic uncertainty for nonlinear system identification while keeping the predictive accuracy of the deterministic version. It learns a Gaussian prior over a low-dimensional latent parameter manifold and an encoder that maps observed data to an approximate posterior; per-task MAP estimation plus Laplace approximation yields a posterior covariance. Across a static sine regression task and the Bouc–Wen hysteretic benchmark, the method matches the deterministic predictor's FIT/RMSE while producing ±3σ intervals whose empirical coverage is close to the nominal 99.7% and whose width shrinks monotonically as more adaptation data arrive. If this holds, practitioners can get both point predictions and trustworthy uncertainty in data-starved identification problems.

Core claim

The central discovery is that the deterministic manifold meta-learning idea—constraining a large neural network's weights to a learned low-dimensional manifold—can be recast in variational form without paying an accuracy penalty. An encoder outputs a Gaussian posterior over the manifold's latent coordinates, trained with an ELBO objective; at test time, the encoder's mean initializes a MAP optimization, and the inverse Hessian of the negative log-posterior provides a closed-form Laplace covariance. On the two benchmarks, predictive performance is comparable to the deterministic baseline, while the resulting uncertainty bands are calibrated in severe low-data regimes and shrink with adaptatio

What carries the argument

The load-bearing object is the learned low-dimensional manifold Pγ(φ), an affine lifting map from n_φ latent coordinates into the n_θ parameter space of an overparameterized base network. A bidirectional-GRU or DeepSets encoder E_ψ maps each dataset to (μ, σ), defining a diagonal Gaussian variational posterior; the meta-training loss is a β-weighted ELBO that balances output reconstruction against KL to a fixed isotropic prior. At test time, the negative log-posterior L_nlp(φ) is minimized from the encoder initialization, and the posterior covariance is the inverse Hessian of this loss at the MAP point. Because n_φ is small, the Hessian inversion is cheap, which is what makes principled unce

Load-bearing premise

The posterior over the low-dimensional latent parameters is assumed to be unimodal enough that one Gaussian centered at the MAP describes it; for hysteretic systems with near-identical loops from different parameter combinations, this can understate uncertainty.

What would settle it

Run the Bouc–Wen adaptation with adaptation lengths much shorter than 100 or with narrowband rather than broadband multisine excitation, repeatedly sample different initialization seeds, and measure the empirical coverage of the ±3σ Laplace intervals on a validation trajectory; if coverage falls substantially below 99.7% in those regimes, the calibration claim does not transfer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Low-data system identification can return calibrated confidence intervals, not just point predictions, making the approach usable in safety-critical settings.
The prior learned over the manifold doubles as a generative model: sampling from N(0,I) and decoding produces plausible system behaviors consistent with the meta-dataset.
Because the posterior covariance is obtained from a low-dimensional Hessian, uncertainty quantification remains computationally affordable even when the base network is large.
The monotonic shrinkage of predictive variance with adaptation length gives a principled, data-driven stopping criterion for data collection.
The equal accuracy with the deterministic version means uncertainty can be obtained at no apparent cost to FIT/RMSE.

Where Pith is reading between the lines

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

The paper's own Remark 3 implies the advertised calibration may be optimistic for very short or narrowband excitations; a natural test is coverage under L<100 or band-limited inputs, where multimodality of the Bouc–Wen posterior is acknowledged.
A mixture-of-Laplace extension, fitting several MAP restarts and combining their covariances, would address the stated unimodality limitation and could turn the method into a multimodal uncertainty estimator.
The generative prior suggests a cheap way to do prior-checking: compare unconditionally sampled trajectories against physical constraints; any mismatch would indicate the meta-dataset prior is not the true generative model.
The same amortized-encoder-plus-MAP-plus-Laplace recipe could be transferred to other meta-learning settings beyond system identification, wherever a low-dimensional task manifold exists.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Worth reading: a clean, self-aware probabilistic extension of the authors' deterministic manifold meta-learning; the Laplace unimodality caveat is real and central, so it is a conditional accept rather than a clear one.

read the letter

The paper is worth knowing about: it grafts a VAE/ELBO training scheme and a MAP + Laplace test-time recipe onto the authors' deterministic manifold meta-learning framework [9]. That combination is new for this niche, and the execution is solid and honest. The ELBO derivation is standard, the reparameterization trick is used correctly, the code is public, and the two benchmarks (sines and Bouc–Wen) are the right tests for the claim. The empirical headline—accuracy comparable to the deterministic baseline plus uncertainty bounds that shrink with data—is supported by the reported numbers, with the caveat that Fig. 4 lacks error bars over the 100 Monte Carlo runs and Fig. 5b has no confidence intervals.

The soft spots are where the reader's stress-test points, and they are real. The advertised "calibrated uncertainty in severely low-data regimes" rests on the Laplace approximation in Eq. (12b). The paper's own Remark 3 concedes that for Bouc–Wen, several (gamma, beta, delta, nu) combinations can produce near-identical force–deformation loops, especially under short or narrowband excitation, and the coverage experiments start at L=100 on the nominal trajectory—so the test does not stress this failure mode. That is not a manufactured concern; it is the central claim's boundary condition. The paper would be materially stronger with a narrowband or short-excitation experiment added. The fact that the authors flag it themselves helps the prose but doesn't fill the gap.

Two more issues, in descending importance. No UQ baselines (deep ensembles, MC dropout, neural processes) are compared, so the relative value of the Laplace uncertainty is unestablished. And the learned sigma_e and beta are load-bearing for the scale of the bounds, yet get no sensitivity analysis. Minor: the abstract says a "generative prior" is learned, but Remark 1 fixes it to N(0,I); that's a wording overreach, not an error.

Verdict: this deserves a serious referee. It is a within-field method paper, not a breakthrough, but it is honestly written, reproducible, and the core idea—amortized variational inference as a probabilistic twist on manifold meta-learning—is reasonable and clearly presented. A reviewer should push on the Laplace unimodality claims, but the paper is in good shape for a solid journal submission after moderate revision. I'd cite it in a related paper and would bring it to a reading group focused on meta-learning or uncertainty quantification.

Referee Report

4 major / 4 minor

Summary. The paper proposes a variational extension of the manifold meta-learning framework for neural system identification. The method trains an encoder to output a diagonal Gaussian over a low-dimensional latent parameter, using a β-VAE-style ELBO with a fixed N(0,I) prior. At meta-test time, the encoder's mean initializes a MAP optimization, and a Laplace approximation at the MAP provides a posterior covariance. The method is evaluated on a synthetic sine regression task and the Bouc–Wen hysteretic benchmark, where it is compared to the deterministic counterpart in terms of predictive accuracy and uncertainty calibration (coverage of ±3σ intervals).

Significance. If the claims hold, the paper offers a simple and computationally feasible way to add uncertainty quantification to a recently proposed meta-learning approach for system identification. The combination of amortized variational inference with test-time MAP+Laplace is sensible, and the paper provides code and detailed experimental settings. However, the headline claim of 'calibrated uncertainty bounds in severely low-data regimes' is not yet fully supported: the quantitative coverage results are limited to a regime the authors themselves identify as unchallenging for the Laplace approximation, and the predictive-accuracy comparison is not uniformly favorable at the smallest data lengths.

major comments (4)

[§IV-B, Fig. 5b and Remark 3] The central claim of calibrated uncertainty in 'severely low-data regimes' is not substantiated by the coverage experiment. Fig. 5b only reports coverage for adaptation lengths L≥100 on the nominal multisine trajectory, while Remark 3 concedes that this does not stress-test the Laplace approximation's failure mode of multimodal posteriors under short or narrowband excitation. Since L=100 is still relatively informative for the Bouc–Wen system, the reported 99.5–99.9% coverage may not transfer to the low-data regime advertised in the abstract. Please either add coverage experiments at shorter L (e.g., 20, 50) or with narrowband inputs, or temper the claim to state that calibration is demonstrated only in the regimes tested.
[§IV-A, Table II] The abstract claims predictive accuracy 'comparable to its deterministic counterpart.' This is not true at N=2 in Table II, where the VAE RMSE is 0.634±0.909 versus 0.548±0.723 for the deterministic method. The difference may not be statistically significant, but the claim should be qualified, and ideally a paired statistical test should be reported. At N=4 and N=10 the VAE is better, so a nuanced statement is appropriate.
[§IV-B, Fig. 5b] The coverage values are reported without confidence intervals. With n_mc=100 adaptation runs, the effective number of independent test points for a 99.7% coverage estimate is important; a small number of outside points can shift the estimate by several tenths of a percent. Please report binomial confidence intervals or the total number of test samples used to compute each coverage point, so the reader can judge whether deviations from 99.7% (e.g., 99.5% at some L) are meaningful.
[§III-B, Eq. (12b)] The Laplace approximation assumes the Hessian at the MAP is positive definite. If the MAP lies near a saddle or the Hessian has negative eigenvalues, the covariance is not valid. The paper does not check or discuss this. Since the loss involves a nonlinear neural network, this is a practical concern; please add a diagnostic (e.g., minimum eigenvalue of the Hessian across test tasks) or discuss why this is not an issue in the experiments.

minor comments (4)

[§III-A, Eq. (8)] In the expression for the evidence, there is an extra 'pγ,ω' in the integral: 'pγ,ω (y(i)|u(i)) = ∫_φ pγ(y(i)|u(i), φ)pω(φ)pγ,ω dφ'. Remove the stray factor.
[§II-B] The sentence 'at meta-test time (Section III-B) it is always followed by gradient-based MAP refinement ... via gradient descent.' contains an orphaned 'via gradient descent.' Please rephrase.
[Table II] The formatting of the first row, 'N= 20.548±0.723', is confusing; it should read 'N=2: 0.548±0.723'.
[§IV-D, Figs. 5–6] Fig. 5a and Fig. 6 do not indicate whether the reported uncertainty is only epistemic (Laplace posterior) or also includes the likelihood noise σ_e. Please clarify how the total predictive variance is computed.

Circularity Check

0 steps flagged

No significant circularity: the variational uncertainty estimates are computed from the task's own negative log-posterior and validated on held-out data; self-citations are baseline, not load-bearing.

full rationale

The derivation chain is self-contained. The ELBO (Eqs. 8 and 10) is standard amortized variational inference; the prior is fixed at N(0,I) per Remark 1. At meta-test time, the MAP estimate (12a) and the Laplace covariance (12b) are computed from the new task's own negative log-posterior loss, with no coverage-calibrating parameter fitted to the reported uncertainty metrics. The Bouc-Wen evaluation reserves the official benchmark as a meta-test set: the reduced-complexity architecture is optimized exclusively on synthetic meta-training data, and adaptation is performed on subsequences of the provided training set, so the coverage in Fig. 5b is an out-of-sample check rather than a fitted re-statement. The only self-citations, [9] and [18], provide the deterministic baseline and the experimental setup; they are not invoked as evidence for the calibrated-uncertainty claim. Remark 3 explicitly concedes that the Laplace approximation can understate uncertainty under short or narrowband excitation and that the coverage study does not stress-test this failure mode; this is an honest correctness limitation, not a circular step, because the reported calibration is still an empirical result over the tested regime rather than an identity. No prediction reduces to its inputs by construction, and no load-bearing uniqueness theorem or ansatz is smuggled in via self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

No truly invented physical entities are introduced. The free parameters that matter for the headline UQ claim are σ²_e and β (plus the manifold dimension), and the key unproven premise is the unimodality of the latent posterior — disclosed by the authors in Remark 3.

free parameters (4)

Likelihood noise variance σ²_e = not reported
Learned together with likelihood parameters γ during meta-training (Sec. III.A); it scales the Hessian in Eq. (12b) and therefore sets the width of the predictive uncertainty bounds whose calibration is the headline claim.
KL weight β = 0.1 (both benchmarks)
Chosen by hand; Sec. III.A.2 says "Tuning β proved critical ... to actively prevent posterior collapse," so the meaningfulness of the latent representation, and hence the UQ, depends on this tuned constant.
Latent manifold dimension n_ϕ = 2 (sines), 20 (Bouc–Wen)
n_ϕ=2 uses oracle knowledge of the true two-factor sine generator; n_ϕ=20 is a design choice for Bouc–Wen. The quality of the UQ and the form of the Laplace posterior depend on this choice.
Network sizes and training schedule = Table I; lr 1e-4→1e-5, 200k iterations, batch 128
Standard hyperparameters, but the FIT and coverage results come from one configuration with no ablations or sensitivity analysis.

axioms (5)

domain assumption All tasks share a low-dimensional latent structure (Eq. 1): p(D) = p(u) ∫ p(z)p(y|u,z) dz.
The manifold hypothesis inherited from [9]; the whole method (n_ϕ ≪ n_θ) assumes it. If systems vary along more factors than the manifold captures, both accuracy and calibration break.
domain assumption Fixed isotropic Gaussian prior N(0,I) over ϕ is expressive enough; the lifting function absorbs mean and covariance.
Remark 1 argues non-restrictiveness. This holds only if P_γ is flexible enough (affine in the experiments, so the induced prior over θ remains Gaussian); it constrains the family of learnable output priors.
ad hoc to paper The Laplace approximation is locally valid (single Gaussian around the MAP, Eq. 12b).
Remark 3 concedes multimodality is plausible for Bouc–Wen and that Fig. 5b does not test it; the calibration claim depends on this approximation.
domain assumption The diagonal-Gaussian amortized encoder E_ψ approximates the true posterior well enough for meta-training.
Standard VI limitation; no evaluation of the encoder's posterior quality against the MAP/Laplace posterior is given.
standard math Unbiased ELBO gradients (reparameterization trick) and validity of the ELBO bound.
Textbook VI theory (Sec. III.A.1); used without proof, appropriate for the venue.

invented entities (1)

Low-dimensional latent parameter ϕ with manifold/lifting function P_γ no independent evidence
purpose: Carries per-task system identity; the object whose posterior is learned and refined.
Inherited from the deterministic manifold framework [9]; a standard statistical latent, not a new physical entity, and it has no falsifiable experimental handle outside the modeling framework itself.

pith-pipeline@v1.3.0-alltime-deepseek · 10256 in / 24353 out tokens · 231884 ms · 2026-08-01T13:49:00.241057+00:00 · methodology

0 comments

read the original abstract

Deep learning has proven highly effective for nonlinear system identification, but heavily parameterized neural networks are prone to overfitting in low-data regimes and lack reliable uncertainty quantification. The recently developed manifold meta-learning framework addresses the data efficiency problem by restricting the model parameters to a meta-learned low-dimensional manifold. However, that method is purely deterministic. We propose a fully probabilistic extension of the manifold meta-learning framework, based on amortized Variational Inference, where a generative prior over the low-dimensional parameter manifold is learned. During task-specific adaptation, we combine Maximum A Posteriori estimation with the Laplace approximation to yield a mathematically grounded posterior approximation. Evaluated on a static regression task and the Bouc--Wen dynamical system benchmark, the proposed approach achieves predictive accuracy comparable to its deterministic counterpart while successfully providing calibrated uncertainty bounds in severely low-data regimes.

Figures

Figures reproduced from arXiv: 2607.18965 by Dario Piga, Marco Forgione, Matteo Rufolo.

**Figure 1.** Figure 1: B. Meta-Testing, MAP and Laplace Approximation At meta-test time, the framework is evaluated on a new, unseen system generating a dataset Dnew = {Dnew tr , Dnew te }. While the encoder Eψˆ provides an immediate, zero-shot Gaussian approximation of the posterior qψˆ(ϕ | Dnew tr ), relying solely on this amortized network can lead to suboptimal accuracy and miscalibrated uncertainty for out-ofdistribution … view at source ↗

**Figure 1.** Figure 1: Overview of the meta-training pipeline. The encoder [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Unconditional generation of 30 trajectories from the prior ϕ ∼ N (0, I). The outputs exhibit the expected sinusoidal pattern characterizing the meta-training distribution. 2) Performance and Uncertainty in Testing: During metatesting, the shared meta-learned representation is used to find the optimal low-dimensional parameters ϕ that fit the training portion of an unseen dataset. This adaptation is imple… view at source ↗

**Figure 3.** Figure 3: Laplace uncertainty bounds on a novel test sequence. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Deterministic and VAE reduced-order models: test [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Performance and uncertainty validation of the varia [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Evolution of model output uncertainty as a function [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 2 linked inside Pith

[1]

Deep networks for system identification: a survey,

G. Pillonetto, A. Aravkin, D. Gedon, L. Ljung, A. H. Ribeiro, and T. B. Sch ¨on, “Deep networks for system identification: a survey,” Automatica, vol. 171, p. 111907, 2025

2025
[2]

Distributionally robust min- imization in meta-learning for system identification,

M. Rufolo, D. Piga, and M. Forgione, “Distributionally robust min- imization in meta-learning for system identification,”IEEE Control Systems Letters, vol. 9, pp. 1652–1657, 2025

2025
[3]

Enhanced trans- former architecture for in-context learning of dynamical systems,

M. Rufolo, D. Piga, G. Maroni, and M. Forgione, “Enhanced trans- former architecture for in-context learning of dynamical systems,” in 2025 European Control Conference (ECC). IEEE, 2025, pp. 819–824

2025
[4]

Can transformers learn optimal filtering for unknown systems?

Z. Du, H. Balim, S. Oymak, and N. Ozay, “Can transformers learn optimal filtering for unknown systems?”IEEE Control Systems Letters, vol. 7, pp. 3525–3530, 2023

2023
[5]

Inverse supervised learning of controller tuning rules,

B. Lakshminarayanan, F. Dett ´u, C. R. Rojas, and S. Formentin, “Inverse supervised learning of controller tuning rules,”Automatica, vol. 178, p. 112356, 2025

2025
[6]

Deep sets,

M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdi- nov, and A. J. Smola, “Deep sets,”Advances in neural information processing systems, vol. 30, 2017

2017
[7]

Set transformer: A framework for attention-based permutation-invariant neural networks,

J. Lee, Y . Lee, J. Kim, A. Kosiorek, S. Choi, and Y . W. Teh, “Set transformer: A framework for attention-based permutation-invariant neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 3744–3753

2019
[8]

Conditional neural processes,

M. Garnelo, D. Rosenbaum, C. Maddison, T. Ramalho, D. Saxton, M. Shanahan, Y . W. Teh, D. Rezende, and S. A. Eslami, “Conditional neural processes,” inInternational conference on machine learning. PMLR, 2018, pp. 1704–1713

2018
[9]

Manifold meta-learning for reduced-complexity neural system identification,

M. Forgione, A. Chakrabarty, D. Piga, M. Rufolo, and A. Bempo- rad, “Manifold meta-learning for reduced-complexity neural system identification,”under revision to Automatica, 2025

2025
[10]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” inInternational Conference on Learning Representations, vol. 2014, 2014

2014
[11]

Tutorial on amortized optimization,

B. Amos, “Tutorial on amortized optimization,”Foundations and Trends in Machine Learning, vol. 16, no. 5, pp. 592–732, 2023

2023
[12]

Variational infer- ence: A review for statisticians,

D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, “Variational infer- ence: A review for statisticians,”Journal of the American statistical Association, vol. 112, no. 518, pp. 859–877, 2017

2017
[13]

Tutorial on variational autoencoders,

C. Doersch, “Tutorial on variational autoencoders,”arXiv preprint arXiv:1606.05908, 2016

Pith/arXiv arXiv 2016
[14]

beta-V AE: Learning basic visual concepts with a constrained variational framework,

I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-V AE: Learning basic visual concepts with a constrained variational framework,” inInternational Conference on Learning Representations, 2017

2017
[15]

Understanding posterior collapse in generative latent variable models,

J. Lucas, G. Tucker, R. B. Grosse, and M. Norouzi, “Understanding posterior collapse in generative latent variable models,” inDeep Generative Models for Highly Structured Data Workshop, ICLR, 2019

2019
[16]

Mixtures of laplace approximations for improved post-hoc uncertainty in deep learning,

R. Eschenhagen, E. Daxberger, P. Hennig, and A. Kristiadi, “Mixtures of laplace approximations for improved post-hoc uncertainty in deep learning,” 2021

2021
[17]

Hysteretic benchmark with a dynamic nonlinearity,

J.-P. Noel and M. Schoukens, “Hysteretic benchmark with a dynamic nonlinearity,” inWorkshop on nonlinear system identification bench- marks, 2016, pp. 7–14

2016
[18]

Meta-learning for physically-constrained neural system identification,

A. Chakrabarty, G. Wichern, V . M. Deshpande, A. P. Vinod, K. Bern- torp, and C. R. Laughman, “Meta-learning for physically-constrained neural system identification,”Neurocomputing, p. 130945, 2025

2025
[19]

Empirical evaluation of gated recurrent neural networks on sequence modeling,

J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

Pith/arXiv arXiv 2014

[1] [1]

Deep networks for system identification: a survey,

G. Pillonetto, A. Aravkin, D. Gedon, L. Ljung, A. H. Ribeiro, and T. B. Sch ¨on, “Deep networks for system identification: a survey,” Automatica, vol. 171, p. 111907, 2025

2025

[2] [2]

Distributionally robust min- imization in meta-learning for system identification,

M. Rufolo, D. Piga, and M. Forgione, “Distributionally robust min- imization in meta-learning for system identification,”IEEE Control Systems Letters, vol. 9, pp. 1652–1657, 2025

2025

[3] [3]

Enhanced trans- former architecture for in-context learning of dynamical systems,

M. Rufolo, D. Piga, G. Maroni, and M. Forgione, “Enhanced trans- former architecture for in-context learning of dynamical systems,” in 2025 European Control Conference (ECC). IEEE, 2025, pp. 819–824

2025

[4] [4]

Can transformers learn optimal filtering for unknown systems?

Z. Du, H. Balim, S. Oymak, and N. Ozay, “Can transformers learn optimal filtering for unknown systems?”IEEE Control Systems Letters, vol. 7, pp. 3525–3530, 2023

2023

[5] [5]

Inverse supervised learning of controller tuning rules,

B. Lakshminarayanan, F. Dett ´u, C. R. Rojas, and S. Formentin, “Inverse supervised learning of controller tuning rules,”Automatica, vol. 178, p. 112356, 2025

2025

[6] [6]

Deep sets,

M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdi- nov, and A. J. Smola, “Deep sets,”Advances in neural information processing systems, vol. 30, 2017

2017

[7] [7]

Set transformer: A framework for attention-based permutation-invariant neural networks,

J. Lee, Y . Lee, J. Kim, A. Kosiorek, S. Choi, and Y . W. Teh, “Set transformer: A framework for attention-based permutation-invariant neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 3744–3753

2019

[8] [8]

Conditional neural processes,

M. Garnelo, D. Rosenbaum, C. Maddison, T. Ramalho, D. Saxton, M. Shanahan, Y . W. Teh, D. Rezende, and S. A. Eslami, “Conditional neural processes,” inInternational conference on machine learning. PMLR, 2018, pp. 1704–1713

2018

[9] [9]

Manifold meta-learning for reduced-complexity neural system identification,

M. Forgione, A. Chakrabarty, D. Piga, M. Rufolo, and A. Bempo- rad, “Manifold meta-learning for reduced-complexity neural system identification,”under revision to Automatica, 2025

2025

[10] [10]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” inInternational Conference on Learning Representations, vol. 2014, 2014

2014

[11] [11]

Tutorial on amortized optimization,

B. Amos, “Tutorial on amortized optimization,”Foundations and Trends in Machine Learning, vol. 16, no. 5, pp. 592–732, 2023

2023

[12] [12]

Variational infer- ence: A review for statisticians,

D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, “Variational infer- ence: A review for statisticians,”Journal of the American statistical Association, vol. 112, no. 518, pp. 859–877, 2017

2017

[13] [13]

Tutorial on variational autoencoders,

C. Doersch, “Tutorial on variational autoencoders,”arXiv preprint arXiv:1606.05908, 2016

Pith/arXiv arXiv 2016

[14] [14]

beta-V AE: Learning basic visual concepts with a constrained variational framework,

I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-V AE: Learning basic visual concepts with a constrained variational framework,” inInternational Conference on Learning Representations, 2017

2017

[15] [15]

Understanding posterior collapse in generative latent variable models,

J. Lucas, G. Tucker, R. B. Grosse, and M. Norouzi, “Understanding posterior collapse in generative latent variable models,” inDeep Generative Models for Highly Structured Data Workshop, ICLR, 2019

2019

[16] [16]

Mixtures of laplace approximations for improved post-hoc uncertainty in deep learning,

R. Eschenhagen, E. Daxberger, P. Hennig, and A. Kristiadi, “Mixtures of laplace approximations for improved post-hoc uncertainty in deep learning,” 2021

2021

[17] [17]

Hysteretic benchmark with a dynamic nonlinearity,

J.-P. Noel and M. Schoukens, “Hysteretic benchmark with a dynamic nonlinearity,” inWorkshop on nonlinear system identification bench- marks, 2016, pp. 7–14

2016

[18] [18]

Meta-learning for physically-constrained neural system identification,

A. Chakrabarty, G. Wichern, V . M. Deshpande, A. P. Vinod, K. Bern- torp, and C. R. Laughman, “Meta-learning for physically-constrained neural system identification,”Neurocomputing, p. 130945, 2025

2025

[19] [19]

Empirical evaluation of gated recurrent neural networks on sequence modeling,

J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

Pith/arXiv arXiv 2014