Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Temporal Variational Implicit Neural Representations

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TV-INRs treats each time series as a time-continuous function and learns a distribution over the weights that generate it, so one trained model can impute and forecast arbitrary new series in a single forward pass.

desk verdict Solid engineering contribution that overclaims its headline low-data result; refereeing should hinge on code, baselines, and toned-down claims. read the letter →

arxiv 2506.01544 v2 pith:3633GO43 submitted 2025-06-02 cs.LG

Batuhan Koyuncu , Rachael DeVries , Ole Winther , Isabel Valera This is my paper
classification cs.LG
keywords TV-INRsimplicitneuralrepresentationstimeseriesimputationforecastingvariationalinferencehypernetworksirregularmultivariateamortized
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

The paper introduces TV-INRs, a probabilistic model that represents irregular multivariate time series as continuous functions of time generated by an MLP whose weights are produced by a hypernetwork. The central claim is that one trained instance can solve many imputation and forecasting tasks at once, without per-sample optimization, by learning an amortized latent distribution over these generator functions. The authors report that this is particularly effective in low-data regimes, where it outperforms existing methods by up to an order of magnitude in imputation MSE, and that it cuts cumulative training time substantially relative to the meta-learning baseline. A sympathetic reader would care because this is a step toward INR-based time series models that are practical for deployment on sparse, irregular data such as clinical records and wearable sensor streams.

What carries the argument

The load-bearing object is the hypernetwork-generated implicit neural representation: $g_\phi([z; c]) = \theta$ defines an MLP $f_\theta$ that maps a Fourier-encoded time coordinate to the feature vector, making the model time-continuous. A Transformer encoder parameterizes both the conditional prior and the approximate posterior over the latent code $z$, and training minimizes the ELBO, the reconstruction of observed and masked values minus the KL divergence between posterior and prior. Masked self-attention keeps the encoder from peeking at values being predicted, and the input representation concatenates the binary mask with zero-filled features and adds a Fourier temporal embedding. This combination is what turns per-sample INR fitting into an amortized, single-forward-pass procedure.

What would settle it

Train TV-INRs on a synthetic square-wave signal with sharp jumps and evaluate imputation MSE within one time step of each discontinuity; if errors there are dramatically larger than elsewhere and the reconstructed curve smooths the jump, the continuity assumption is falsified.

Watch

Extended reading notes

Core claim

TV-INRs couples a conditional prior $p_{\psi_z}(z \mid Y_{\text{obs}}, T_{\text{obs}})$ and an approximate posterior $q_{\gamma_z}(z \mid Y, T)$, both parameterized by a Transformer encoder, with a hypernetwork $g_\phi$ that maps the latent code $z$ (plus static covariates $c$) to the weights $\theta$ of an INR $f_\theta$. The generative model factorizes as $p(Y_m, Y_{\text{obs}}, z \mid Y_{\text{obs}}, T, c) = p_{\psi_z}(z\mid Y_{\text{obs}},T_{\text{obs}}) \prod_{l} p_{\theta(z,c)}(y_l \mid t_l)$, so after training, missing stamps and future horizons are predicted by encoding the observed context and evaluating $f_\theta$ at the requested time coordinates. The paper reports that this single instance matches or beats TimeFlow and SAITS on imputation across observation rates 5%, 30%, and 50%, and DeepTime and TimeFlow on forecasting horizons 96, 192, 336, and 720, with the largest advantages at short signal lengths ($L=200$), e.g., Solar-10 imputation MSE of 0.038 versus 0.330 for TimeFlow and 0.066 for SAITS. Throughout, the method requires no fine-tuning or meta-learning at test time, only a forward pass.

Load-bearing premise

The model assumes each time series is a smooth function of time that an MLP can generate, so sharp discontinuities or mixed discrete-continuous channels must be absorbed by the latent code.

Editorial extensions

If this is right

  • One TV-INRs instance covers imputation at 5%, 30%, and 50% observation and forecasting at horizons 96, 192, 336, and 720, where baselines need a separately trained model per setting.
  • Test-time cost does not grow with the number of test samples or horizon, because prediction is a single forward pass rather than gradient-based meta-learning.
  • In low-data regimes, imputation MSE drops by up to an order of magnitude relative to TimeFlow, for example on Solar-10 at $L=200$.
  • Cumulative training time is 1.30–2.81× lower than TimeFlow for imputation and 2.41–3.70× lower for forecasting.
  • Conditioning the generator on covariates (C-TV-INRs) gives the largest accuracy gains under extreme sparsity, such as P12 at $\tau=0.10$ and HAR at 95% missingness.

Reading between the lines

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

  • Beyond the paper, the same architecture could be read as a distribution over functions, so drawing several latent samples $z$ would yield a predictive distribution for each missing value; the paper does not report uncertainty calibration.
  • Because the hypernetwork outputs all INR weights, memory scales with the generator's parameter count; on longer or higher-dimensional series, low-rank or per-sample positional-embedding variants are the natural next test.
  • The continuity assumption suggests a direct extension: add a discrete regime-switch variable to the latent code to handle abrupt changepoints, and compare against TV-INRs on piecewise-constant signals.
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

4 major / 5 minor

Summary. The paper introduces Temporal Variational Implicit Neural Representations (TV-INRs), a probabilistic model in which a Transformer encoder maps observed, irregularly sampled multivariate time series (and optional covariates) to a latent code z; a hypernetwork then generates the weights of an implicit neural representation (INR) so that a single forward pass produces imputations or forecasts at arbitrary time points. The authors compare TV-INRs against TimeFlow, SAITS, and DeepTime on Electricity, Traffic, Solar, HAR, and P12, and report that TV-INRs is particularly effective in low-data regimes, including order-of-magnitude MSE improvements for imputation, while being more computationally efficient because one model handles multiple missingness rates and forecast horizons without per-sample fine-tuning.

Significance. If the empirical claims are substantiated, the paper would make a useful contribution: amortized variational inference over INR weights is a plausible route to individualized and scenario-agnostic time series modeling, and the reported computational savings over per-scenario training are practically relevant. The paper also provides a probabilistic treatment that could enable uncertainty quantification, which is a strength. However, the central quantitative claim is currently supported by a single, high-variance experimental cell, and several reporting and notation issues undermine reproducibility. The method itself is not obviously flawed, but the evidence presented does not yet support the strength of the headline claims.

major comments (4)
  1. [Abstract; Section 4.1.1; Table 2] The abstract's claim that TV-INRs 'outperforms existing methods by an order of magnitude in mean squared error for imputation task' rests on a single cell in Table 2: Solar-10, L=200, τTest=0.50, where TV-INRs reports 0.038±0.031 versus TimeFlow's 0.330±0.046 and SAITS's 0.066±0.035. The ratio to TimeFlow is approximately 8.7, not 10; the TV-INRs error interval (0.007–0.069) overlaps substantially with SAITS's (0.031–0.101); and the coefficient of variation is 81%. With only three seeds, the reported p<0.05 significance is not supported by the displayed means and standard deviations, and no per-seed results or test details are given. Moreover, other low-data cells in the same table show only 1.3–3x improvements, so 'several datasets' and 'order of magnitude' overstate the evidence. Please provide per-seed results, the exact statistical test and p-values, and either strengthen the evidence or temper the claims.
  2. [Table 2] Table 2 is severely malformed: the Solar-10 results are interleaved with the L dimension in a way that makes the row/column alignment unclear, e.g., under the SAITS 2K row the entry '10K 1.086 ± 0.005' appears without a clear column header, and the same happens for the L=200 block. This makes the key low-data results difficult to read and reproduces poorly. The table should be restructured, for instance with a separate row or grouped sub-table for each signal length.
  3. [Section 3.1, Eq. (2); Section 4, Imputation task] Equation (2) is not a well-formed generative model: the left-hand side conditions on Yobs while also listing Yobs as a generated variable; the likelihood is written as pθ(z,c)(yl|tl), which uses the random variables z and c as subscripts rather than defining the generated parameters θ; and the product over l includes the observed entries, contradicting the conditioning. Please rewrite the joint distribution as, e.g., p(Ym, z | Yobs, T, c) = pψz(z | Yobs, Tobs) ∏_{l∈M} p(yl | tl, θ(z,c)), or an equivalent explicit form. In addition, Section 4 states that imputation uses the 'approximate posterior distribution pψz(z|Yobs,Tobs)', but pψz is introduced in Section 3.1 as the conditional prior; the approximate posterior used during training is qγz. Since the test-time procedure is to use the conditional prior in place of the posterior, this naming inconsistency should be resolved and the inference step clearly explained.
  4. [Section 4, multivariate datasets paragraph] The statement 'TV-INRs is the first temporal INR model to handle such multivariate signals' is unsupported: no literature search or citation is provided, and it is used to justify excluding TimeFlow from the multivariate comparisons in Table 4. Without a fair INR-based multivariate baseline, the multivariate imputation results in Table 4 (which show TV-INRs outperforming only SAITS) do not support the 'first' claim. Please either cite prior multivariate INR time series work or drop the 'first' claim and, if possible, add a multivariate INR baseline.
minor comments (5)
  1. [Section 3.2, Eq. (4)] The set U(i) is used in Eq. (4) to denote entries that are filled with zeros, but U(i) is never defined; earlier the paper defines O(i), M(i), and A(i). Please define U(i) or use the already-defined sets consistently.
  2. [Section 2.1] There is a typo: 'continous' should be 'continuous'.
  3. [Appendix 6.5, Table 8; Appendix 6.8, Table 13] The text and Table 4 specify P12 with L=48, but Table 13 lists 'TV-INR P12 128' and Table 8 does not list P12 hyperparameters at all. Please align the P12 length and provide the corresponding hyperparameter settings.
  4. [Section 4.1.3 and Appendix 6.8] The training-time comparison is presented as a ratio of cumulative training times (TV-INRs vs. TimeFlow), but the reader is not told whether the reported TV-INRs training time includes all hyperparameter search runs or only the final model; please clarify the protocol for measuring training time.
  5. [Section 4.1.1, Table 2] The notation 'TV-INRs 2K ∼ S' and 'TV-INRs 200 ∼ S' is not explained; the tilde and 'S' presumably indicate the stochastic τTrain scheme, but this should be defined in the caption or text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TV-INRs presents empirical benchmark claims under a standard amortized variational inference protocol; existing self-citations are background material, not load-bearing.

full rationale

The paper does not derive a first-principles result that could collapse into its own inputs; its central claims are empirical. TV-INRs is trained by maximizing the ELBO in Eq. (3), using observed and artificially masked entries, and then evaluated on held-out masked positions as specified in Eqs. (10)-(13). No fitted parameter is renamed as a prediction, and no reported MSE/MAE value equals a training target by construction. The only self-citations (Koyuncu et al., 2023; Peis et al., 2025) appear in the background section as examples of hypernetworks and INR parameter generation; they are not used to justify the model choice or to forbid alternatives, and no uniqueness theorem is imported. The headline order-of-magnitude claim rests on a single high-variance Solar-10 cell in Table 2, but statistical fragility is a robustness/correctness concern, not circularity. Eq. (2) has a conditioning typo in which Yobs appears on both sides of the joint distribution, but the implemented training objective and experimental protocol clearly condition the decoder on the latent code and timestamps, so no reported result is true by definition. Overall, the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The model is a deep learning method, so the free parameters are hyperparameters chosen by grid search and reported in the appendix. The axioms are the modeling assumptions behind the VAE-INR architecture, especially the continuity of time series and the representativeness of self-supervised masking. No new physical entities are introduced; the latent z is a standard statistical latent variable.

free parameters (5)
  • Latent dimension dim_z = 32 or 64 depending on dataset
    Selected by grid search over {16,32,64} (Appendix Table 10); controls capacity of the global code z and is central to the model's individualized representation.
  • Random Fourier Features dimension m and scale sigma = m=256, sigma=2 for all experiments
    Chosen from grid search over m in {128,256,512}, sigma in {1,2,4}; controls the temporal frequency resolution of the INR input encoding.
  • Hypernetwork and generator widths = Hypernetwork [128,256]; generator [64,64,64] or [64,64,64,64]
    Chosen from grid search; determine the capacity of the weight-generating and signal-generating MLPs.
  • Transformer encoder size = d_model=128, heads 2 or 4, layers 2 or 4
    Chosen from grid search over d_model {64,128,256}, layers {2,4,6}, heads {2,4,8}; affects how well the encoder summarizes observed values.
  • Learning rate = 1e-4
    Grid searched over {1e-5,1e-4,5e-4}; standard training hyperparameter.
assumptions (5)
  • standard math Variational inference with the standard ELBO is a valid way to train the model.
    Used in Eq. (3); relies on Kingma and Rezende's VAE framework.
  • domain assumption Each time series can be represented as a continuous function of time with bounded complexity, so an MLP with hypernetwork-generated weights can fit or extrapolate it.
    The generator f_theta(t_l)=y_l in Eq. (2) and Section 3.1 assumes continuity; irregular or discontinuous clinical signals may violate this.
  • domain assumption The latent z and covariates c contain all information needed to generate every channel at every time point.
    The generative model conditions only on z and c via theta = g_phi(z,c); any within-signal variation not captured by z,c must be picked up by the INR itself, which is deterministic given theta.
  • domain assumption Artificially masking observed values during training produces missingness patterns representative of real test-time missingness (missing at random).
    Section 3.2 masks random observed entries; if real missingness is informative (e.g., sicker patients missing more data), the model may be biased.
  • domain assumption Transformer encoders can summarize arbitrary-length irregular sequences into a fixed-size latent distribution.
    The approximate posterior and conditional prior in Sections 3.1-3.2 rely on self-attention over variable-length inputs.
invented entities (1)
  • Latent variable z (global time-series code)
    purpose: Condenses an entire signal into a vector that the hypernetwork uses to generate INR weights, enabling individualization and forecasting.
    A standard VAE latent construct; the paper provides no external falsifiable prediction for z itself, only downstream task performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temporal Variational Implicit Neural Representations." pith.science (2026). https://pith.science/paper/3633GO43

@misc{pith2026250601544,
  author       = {Pith},
  title        = {Pith review of: Temporal Variational Implicit Neural Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3633GO43}},
  note         = {Machine review of arXiv:2506.01544}
}
read the original abstract

We introduce Temporal Variational Implicit Neural Representations (TV-INRs), a probabilistic framework for modeling irregular multivariate time series that enables efficient and accurate individualized imputation and forecasting. By integrating implicit neural representations with latent variable models, TV-INRs learn distributions over time-continuous generator functions conditioned on signal-specific covariates. Unlike existing INR approaches that require extensive training, fine-tuning or meta-learning, our method achieves accurate individualized predictions through a single forward pass. Our experiments demonstrate that with a single TV-INRs instance, we can accurately solve diverse imputation and forecasting tasks, offering a computationally efficient and scalable solution for real-world applications. TV-INRs performs particularly well in low-data regimes, where on several datasets it achieves substantially lower imputation error, including order-of-magnitude improvements.

Figures

Figures reproduced from arXiv: 2506.01544 by the authors.

Figure 1
Figure 1. Visualization of temporal stamps T , features Y , mask Ω, and static covariates C. T and Y represent the input signal, Ω indicates missing values with binary entries, and C contains time-invariant covariates. for this task [Chen et al., 2024]. Moreover, they may have trouble identifying and preserving key information when attending to large inputs [Wen et al., 2022]. Recently, INRs have been used in continuous model… view at source ↗
Figure 2
Figure 2. Graphical models for generative and inference tasks. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Classification performance at various levels of missingness, as measured by AUC-ROC. A higher AUC [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: TV-INRs imputation predictions for Electricity dataset [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: TV-INRs imputation predictions for Electricity dataset [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: TV-INRs forecasting predictions for Traffic dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: TV-INRs imputations for HAR dataset. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 13 canonical work pages

  1. [3]

    Gated feedback recurrent neural networks

    Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent neural networks. In International conference on machine learning, pages 2067–2075. PMLR,

    work page

  2. [4]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long

    URL https://arxiv.org/abs/2103.01600. Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625,

    arXiv

  3. [7]

    Moreover, training data precedes validation data

    For univariate datasets, test windows are extracted sequentially from the end of each time series. Moreover, training data precedes validation data. 2NO: Non-overlapping, FE: From end of the series 12 Temporal Variational Implicit Neural Representations Table 5: Dataset splitting details for univariate imputation experiments. Training and validation sets ...

    work page 2000

  4. [9]

    doi:10.1007/978-3-031-16788-1_20. M. Niemeyer, J. T. Barron, B. Mildenhall, M. S. M. Sajjadi, A. Geiger, and N. Radwan. Regnerf: regularizing neural radiance fields for view synthesis from sparse inputs. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

    work page doi:10.1007/978-3-031-16788-1_20 2022

  5. [10]

    Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Goli ´nski, Yee Whye Teh, and Arnaud Doucet

    doi:10.1109/cvpr52688.2022.00540. Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Goli ´nski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022b. Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In ...

    arXiv 2022

  6. [12]

    10 Temporal Variational Implicit Neural Representations Tom Bamford, Elizabeth Fons, Yousef El-Laham, and Svitlana Vyetrenko

    URL https://arxiv.org/abs/2201.11950. 10 Temporal Variational Implicit Neural Representations Tom Bamford, Elizabeth Fons, Yousef El-Laham, and Svitlana Vyetrenko. Mads: Modulated auto-decoding siren for time series imputation. arXiv preprint arXiv:2307.00868,

    arXiv

  7. [13]

    Robust online time series prediction with recurrent neural networks

    Tian Guo, Zhao Xu, Xin Yao, Haifeng Chen, Karl Aberer, and Koichi Funaya. Robust online time series prediction with recurrent neural networks. In 2016 IEEE international conference on data science and advanced analytics (DSAA), pages 816–825. Ieee,

    work page 2016

  8. [14]

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun

    URL https://arxiv.org/abs/2211.14730. Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. Transformers in time series: A survey. arXiv preprint arXiv:2202.07125,

    arXiv

Show all 22 references
  1. [15]

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi

    URL https://arxiv.org/abs/2208.05836. Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. Learning deep time-index models for time series forecasting. In International Conference on Machine Learning, pages 37217–37237. PMLR,

    arXiv

  2. [16]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

    arXiv

  3. [17]

    ISBN 9781450342322

    Association for Computing Machinery. ISBN 9781450342322. doi:10.1145/2939672.2939785. URL https://doi.org/10.1145/2939672.2939785. Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In European Conference on Computer Vision, pages ...

  4. [18]

    11 Temporal Variational Implicit Neural Representations 6 Appendix A 6.1 Datasets In this section, we provide more details about the datasets we have used. We start with the list of uni-variate datasets: Electricity Dataset records hourly electricity consumption from 321 custo...

    2012

  5. [19]

    15 Temporal Variational Implicit Neural Representations Table 13: Training times for imputation task, TV-INRs. Model Name Dataset L Max Epochs Training Time TV-INR Electricity 200 2000 8h45m TV-INR Electricity 2000 4000 12h55m TV-INR Traffic 200 2000 10h35m TV-INR Traffic 2000...

  6. [21]

    draw_ratio

    Table 8: Hyperparameter details of TV-INRs for imputation task. ELECTRICITY TRAFFIC SOLAR-10 HAR L 200 2000 200 2000 200 10000 128 dim_z 32 64 32 64 32 64 32 epochs 2000 4000 2000 4000 2000 4000 3000 bs 256 64 256 64 256 32 128 lr 1e-4 1e-4 1e-4 1e-4 1e-4 1e-4 1e-4 Transformer...

    2000

  7. [2006]

    Additionally, there is an hourly version of this dataset, known as Solar-Hourly. For some datasets, the feature vectors Y (i) = {y(i) l }Li l=1 expand from univariate ( d = 1 ) to multivariate ( d > 1), with each dimension representing a unique sensor used to collect observati...

    2012

  8. [2012]

    In 2012 computing in cardiology, pages 245–248. IEEE,

    2012

  9. [2016]

    Meta-learning via hypernetworks

    Dominic Zhao, Seijin Kobayashi, João Sacramento, and Johannes von Oswald. Meta-learning via hypernetworks. In 4th Workshop on Meta-Learning at NeurIPS 2020 (MetaLearn 2020). NeurIPS,

    2020

  10. [2021]

    Batuhan Koyuncu, Pablo Sanchez-Martin, Ignacio Peis, Pablo M Olmos, and Isabel Valera

    URL https://arxiv.org/abs/2102.04776. Batuhan Koyuncu, Pablo Sanchez-Martin, Ignacio Peis, Pablo M Olmos, and Isabel Valera. Variational mixture of hypergenerators for learning distributions over functions. arXiv preprint arXiv:2302.06223,

    arXiv

  11. [2022]

    Emilien Dupont, Yee Whye Teh, and Arnaud Doucet

    URL https://arxiv.org/abs/2005.08482. Emilien Dupont, Yee Whye Teh, and Arnaud Doucet. Generative models as distributions of functions. CoRR, abs/2102.04776,

    2005 arXiv

  12. [2023]

    doi:https://doi.org/10.1016/j.ecosta.2022.02.001

    ISSN 2452-3062. doi:https://doi.org/10.1016/j.ecosta.2022.02.001. URL https://www.sciencedirect. com/science/article/pii/S2452306222000090. 9 Temporal Variational Implicit Neural Representations Ikaro Silva, George Moody, Daniel J Scott, Leo A Celi, and Roger G Mark. Predictin...

    2022 doi

  13. [2024]

    David Ha, Andrew Dai, and Quoc V Le

    URL https://openreview.net/forum?id=FpElWzxzu4. David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106,

    arXiv

  14. [2025]

    Chelsea Finn and Sergey Levine

    URL https://arxiv.org/abs/2504.16580. Chelsea Finn and Sergey Levine. Meta-learning and universality: Deep representations and gradient descent can approximate any learning algorithm. arXiv preprint arXiv:1710.11622,

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.