Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read MDIR claims that LLM weight homology can be detected with perfect accuracy on LeaFBench by comparing a single pair of weight matrices and computing a large-deviation p-value, with no model inference.

desk verdict Serious, novel fingerprinting method for LLM weight homology, marred by an abstract that belongs to a different paper and p-value machinery that is shakier than the prose. read the letter →

arxiv 2508.06301 v1 pith:6NRXYVPW submitted 2025-08-08 cs.LG cs.AIcs.CVcs.DC

Junhyeog Yun , Minui Hong , Gunhee Kim This is my paper
classification cs.LGcs.AIcs.CVcs.DC
keywords modelhomologyweightprovenancelargedeviationtheorypolardecompositionHaarmeasurep-valueLeaFBenchLLMfingerprinting
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 argues that whether one large language model's weights were derived from another's can be answered by looking at a single pair of matrices, with no inference or probing. The proposed method, MDIR, aligns two embedding matrices via the orthogonal factor of the polar decomposition, finds the best permutation with the Hungarian algorithm, and converts the alignment trace into a p-value using large deviation theory. On the LeaFBench benchmark MDIR reports perfect AUC and accuracy, and it reconstructs layer correspondences for pruned and upscaled models. If MDIR is right, model provenance checks become cheap, statistical, and robust to changes such as tokenizer replacement and heavy noise.

What carries the argument

The central object is $\tilde{U} = \operatorname{Ortho}(E'^\top E)$, the orthogonal factor in the polar decomposition of the cross-covariance between two embedding matrices (or between shared-token submatrices). Under homology $\tilde{U}$ is close to a permutation, so the maximized trace $\max_P \operatorname{Tr}(P \tilde{U}^\top)$ measures alignment. Large deviation theory for the Circular Real Ensemble (a Haar-random matrix on $O(n)$ / $SO(2m)$) supplies the $p$-value bound $p \le n! \exp(-c^2/2)$ for a trace value $c$, with the Bonferroni correction for $n!$ permutations.

What would settle it

Take a set of, say, 50 pairs of LLMs trained from scratch on the same data and tokenizer but from different random seeds (no shared initialization) and run MDIR's embedding-level test; if a substantial fraction of unrelated pairs yield $p$-values below $10^{-6}$, the Haar-uniform null does not hold for real embeddings and the reported perfect scores would not transfer to this setting.

Watch

Extended reading notes

Core claim

The central claim is that the orthogonal part of the polar decomposition of $E'^\top E$, after matching the optimal permutation, carries enough signal to distinguish homologous from unrelated LLM weights, and that large deviation theory for the orthogonal group converts that signal into a rigorous $p$-value. The paper further claims that this $p$-value is valid across different tokenizers (by restricting to shared tokens), different layer counts (via layer correspondence matching), and common adaptations (fine-tuning, continual pretraining, upcycling, pruning, orthogonal transformations). MDIR reports perfect $1.000$ AUC and accuracy on LeaFBench, making it the first method to achieve this.

Load-bearing premise

The $p$-value calculations assume that when two models are unrelated, the best rotation aligning their embeddings is uniformly random over the orthogonal group under the Haar measure, a claim proven only for i.i.d. Gaussian embeddings, not for real trained anisotropic ones.

Editorial extensions

If this is right

  • MDIR can flag fine-tuned, continually pretrained, upcycled, pruned, or orthogonally transformed models using only weight matrices, on devices with limited compute.
  • Because it outputs genuine p-values rather than uncalibrated similarity scores, a significance threshold can be fixed a priori without tuning on known positive/negative pairs.
  • It reconstructs the correspondence between layers, revealing specific pruning or upscaling strategies, such as which layers of a base model were kept.
  • It remains effective when models use different tokenizers or different numbers of layers, and it degrades only after noise has already destroyed model functionality.

Reading between the lines

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

  • The paper's own ablation uses only 4 independently seeded models as negative pairs; a larger battery of such pairs would be needed to confirm that the Haar-uniform null holds for real trained embeddings, whose statistics differ from the Gaussian case proven in Theorem 2.
  • The large-deviation rate function $I(r)=2r^2$ for $r\le 1/2$ hints at a sharp transition near $r=1/2$, which could be exploited to design a more powerful test statistic or to understand how much obfuscation an adversary can apply before detection fails.
  • The layer-correspondence output could naturally feed a phylogenetic analysis of model families, a direction the paper explicitly leaves open.
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

3 major / 6 minor

Summary. The manuscript (arXiv:2508.06309v3) describes MDIR, a white-box method for detecting and reconstructing weight homology between LLMs. It computes the orthogonal factor Ũ of E'^T E over shared vocabulary embeddings, applies Hungarian matching to estimate a permutation, and converts the maximum trace c = max_P Tr(PŨ^T) into a p-value via a large-deviation rate function. It claims perfect accuracy/AUC on LeaFBench, layerwise mapping for depth-scaled and pruned models, robustness to noise and tokenizer changes, and clean negatives for independently seeded models. Note: the opening abstract/header in the supplied text is for a different paper (FedMeNF); the substantive content is the MDIR paper.

Significance. If the p-value claim holds, MDIR is a significant advance: it needs no model inference, operates on a single matrix pair, handles different tokenizers and layer counts, and outputs interpretable permutation/weight correspondence maps. The empirical package is strong: perfect LeaFBench scores, SOLAR/Llama layer mappings consistent with external technical reports, a controlled changed-tokenizer pretraining experiment, and a seed ablation with clean negatives. The a priori threshold setting is a plus. The main gap is the statistical null model: the Haar-uniform assumption for Ũ is proven only for Gaussian embeddings, and the reported p-values are asymptotic point estimates rather than rigorous upper bounds because the LDT prefactor K is dropped.

major comments (3)
  1. [§3.3, Appendix D, Theorem 2] The p-value null is that Ũ = Ortho(E'^T E) is Haar-uniform on O(n). Theorem 2 proves this only when E' has i.i.d. Gaussian entries. Real LLM embeddings are trained, anisotropic, bf16-quantized, and share token statistics; under the null of independent models, E'^T E may have a systematic non-isotropic component, making Ũ stochastically closer to identity and inflating c = max_P Tr(PŨ^T). Then every reported p-value is understated. Section 4.5 supplies only four seed-trained negative models, and Figure 1/Appendix I only a handful of unrelated pairs—far too few to validate tail probabilities around 10^{-1000} or smaller. This directly undermines the claim of rigorous false-positive control. The block-wise patterns reported in Section 4.5 for inner attention matrices even for different seeds is direct evidence that the Haar null is violated for those matrices.
  2. [§3.3, Algorithm 1] The derivation gives f(c) ≤ K exp(-c^2/2) with an unspecified K, then 'in practice, we set ε=0'. Algorithm 1 uses p = d! exp(-Tr(PŨ^T)^2/2), i.e., K=1. Thus the reported p-values are not rigorous upper bounds; they are asymptotic point estimates modulo a subexponential prefactor. The conclusion's statement that extreme p-values are 'an accurate assessment' overstates the theory. Because -c^2/2 dominates log(n!), classification is robust, but the paper should either calibrate K empirically or report p-values as heuristic significance scores.
  3. [§H, Theorem 3 proof] The LDT rate-function derivation is a plausible variational calculation, but it assumes absolute continuity of the constrained equilibrium measure and uses a Chebyshev expansion without a fully rigorous justification of the constrained infimum, and the exact rate for r > 1/2 is explicitly conjectural. If the p-value machinery is presented as 'rigorous', this part should either be tightened or clearly marked as a heuristic asymptotic approximation.
minor comments (6)
  1. [Abstract/header] The supplied text opens with an abstract about FedMeNF and neural fields, but the full paper is about MDIR for LLM weight homology. This mismatch must be fixed before resubmission.
  2. [References] All reference URLs appear as placeholder strings (e.g., 'URL ...'); the reference list is not in publishable form.
  3. [§4.1] The sentence 'the perfect score of Area-Under-Curve suggests a clear separation threshold between true positives and false negatives' should read '... between positives and negatives' or 'true positives and true negatives'.
  4. [Figure 1] The figure labels are illegible in the provided text; a high-resolution version with readable model names and a color map is needed.
  5. [Algorithm 1 / Appendix F] The layer-significance threshold p0 appears in Algorithm 1 and Appendix F but is never given a concrete value. The paper says a broad range of thresholds works, but the parameter should still be specified for reproducibility.
  6. [§4.4] The noise-injection experiment is a useful stress test, but the statement that an adversary 'would also significantly degrade model performance' is demonstrated only for i.i.d. Gaussian perturbation, not for structured obfuscation; the paper should phrase this more narrowly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; derivation is self-contained, with minor self-citation and an unproven-but-not-circular Haar null caveat.

full rationale

The claimed derivation—compute U-tilde = Ortho(E^T E'), select P = argmax Tr(P U-tilde^T), and report p = n! exp(-c^2/2) with an a priori threshold—does not fit any parameter to homology labels or to benchmark positives/negatives; the LDT rate function is derived in Appendix H, and validation uses external ground truth (LeaFBench, SOLAR, Llama pruning, Qwen upcycling, RWKV lineage). The one step that resembles circularity is the Section 3.3 null assumption that U-tilde is Haar-uniform; however, this is an assumption about the distribution under the null, not a definition of homology, and Theorem 2's Gaussian proof is a domain-support gap rather than a reduction of the conclusion into the input. The p-value is therefore not forced by construction, and the significance threshold is set a priori without post-hoc calibration. The paper contains a minor self-citation (RWKV-5/7 lineage, Peng et al. 2024/2025, co-authored by the present authors), but it is not load-bearing because the lineage is public model documentation and the method's validity does not rest on it. Hence score 2: no significant circularity, with minor caveats around the unproven Haar assumption and limited negative-pair validation.

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

The method is unusually parameter-light: no hand-fitted constants enter the detection statistic, and the significance threshold is a priori. The load-bearing content is concentrated in domain assumptions (validity of the E' = E U^T + N_E model, preservation of invariant coordinates through training, alignment of shared tokens after tokenizer changes) and in the statistical null (Haar uniformity), which is proven only for Gaussian random embeddings. Everything else is standard matrix analysis and random matrix theory. No new physical entities are postulated.

free parameters (2)
  • LDT slack constant epsilon (with implicit K) = 0
    In Section 3.3 the bound is log p <= log(n!) - c^2/2 + epsilon; the authors set epsilon = 0 'for simplicity', dropping the unknown multiplicative constant K from f(c) <= K exp(-c^2/2). Setting epsilon = 0 makes reported p-values more significant than the proven bound strictly allows, though K is subdominant for large n.
  • Layer-significance threshold p_0 = not specified (a priori; paper states 0.001 to 1e-6 gives same classification)
    Algorithms 2 and 3 in Appendix G use 'p < p_0' to decide when a permutation is reliably identified. The paper states any reasonable threshold yields the same LeaFBench classification, so this is not fitted to data, but it is a hand-set constant.
assumptions (6)
  • domain assumption Under the null (unrelated models), Ortho(E'^T E) is Haar-uniform on O(n).
    Invoked in Section 3.3 to convert the trace into a p-value. Appendix D (Theorem 2) proves this only when E' has i.i.d. Gaussian rows; real LLM embeddings are trained and structured, so the transfer of the Gaussian proof to real models is unproven.
  • domain assumption Homologous models satisfy E' = E U^T + N_E with U orthogonal and N_E a modest perturbation.
    Section 3.2 sets up the alignment problem this way. If continued pretraining destroys the shared global orthogonal relation, the trace statistic loses power. The changed-tokenizer experiment shows low trace at 1.64B tokens before alignment strengthens.
  • domain assumption Training trajectories preserve the component of weights along the totally invariant group G.
    Key Insight in Section 2.1. Requires G-invariant optimizers and infinite precision; the paper admits AdamW is not fully orthogonal-invariant, so preservation is approximate and not proven.
  • domain assumption For different tokenizers, embeddings of shared tokens remain approximately aligned up to the global transformation U.
    Section 3.2 'Changed Tokenizer' estimates U using only common tokens. Validated by one Qwen2.5 retraining experiment, but not generally proven.
  • standard math The LDT/random-matrix analysis: n = 2m even, det(U-tilde) = 1, negligible external field, rate I(r) = 2r^2 for r <= 1/2.
    Appendix H: Theorem 3 plus the 'without loss of generality' assumptions, citing Eichelsbacher et al. 2011, Mehta 2004, and the Chebyshev expansion. The exact branch for r > 1/2 is stated as a conjecture (H.3).
  • ad hoc to paper In the equilibrium-measure calculation, the single-particle external field is negligible in the thermodynamic limit.
    Appendix H.1 uses w_m(x) -> 1 to reduce the problem to the interaction term, which lets the Chebyshev computation close. It is plausible (O(m) versus O(m^2) energy) but is a simplification introduced to make the derivation work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields." pith.science (2026). https://pith.science/paper/6NRXYVPW

@misc{pith2026250806301,
  author       = {Pith},
  title        = {Pith review of: FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NRXYVPW}},
  note         = {Machine review of arXiv:2508.06301}
}
read the original abstract

Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. FinInvest-GTCN: Explainable Graph-Temporal-Causal Modeling for Risk-Aware Investment Decision Optimization

    cs.CL 2026-06 unverdicted novelty 4.0 of 10

    FinInvest-GTCN combines graph, temporal, and causal networks with meta-causal adaptation to improve risk-adjusted predictions for VC investments, achieving RA-MSE of 2.51 and 18.7% higher simulated returns on propriet...

Reference graph

Works this paper leans on

72 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Deep learning with differential privacy

    Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In ACM CCS, 2016

    work page 2016

  3. [3]

    A survey on federated learning: The journey from centralized to distributed on-site learning and beyond

    Sawsan AbdulRahman, Hanine Tout, Hakima Ould-Slimane, Azzam Mourad, Chamseddine Talhi, and Mohsen Guizani. A survey on federated learning: The journey from centralized to distributed on-site learning and beyond. IEEE Internet of Things Journal, 2020

    work page 2020

  4. [4]

    Canonical fields: Self-supervised learning of pose-canonicalized neural fields

    Rohith Agaram, Shaurya Dewan, Rahul Sajnani, Adrien Poulenard, Madhava Krishna, and Srinath Sridhar. Canonical fields: Self-supervised learning of pose-canonicalized neural fields. In CVPR, 2023

    work page 2023

  5. [5]

    Federated learning: A survey on enabling technologies, protocols, and applications

    Mohammed Aledhari, Rehma Razzak, Reza M Parizi, and Fahad Saeed. Federated learning: A survey on enabling technologies, protocols, and applications. IEEE Access, 2020

    work page 2020

  6. [6]

    Concentrated differential privacy: Simplifications, extensions, and lower bounds

    Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography, 2016

    work page 2016

  7. [7]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015

    arXiv 2015

  8. [8]

    pfl-bench: A comprehensive benchmark for personalized federated learning

    Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. NeurIPS, 2022 a

    work page 2022

Show all 72 references
  1. [9]

    Federated meta-learning with fast convergence and efficient communication

    Fei Chen, Mi Luo, Zhenhua Dong, Zhenguo Li, and Xiuqiang He. Federated meta-learning with fast convergence and efficient communication. arXiv:1802.07876, 2018

    2018 arXiv

  2. [10]

    Transformers as meta-learners for implicit neural representations

    Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In ECCV, 2022

    2022

  3. [11]

    Learning continuous image representation with local implicit image function

    Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, 2021

    2021

  4. [12]

    Videoinr: Learning video implicit neural representation for continuous space-time super-resolution

    Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vidit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022 b

    2022

  5. [13]

    Neural fields: theory and applications

    Stephen Coombes, Peter beim Graben, Roland Potthast, and James Wright. Neural fields: theory and applications. Springer, 2014

    2014

  6. [14]

    Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024

    Moming Duan, Qinbin Li, Linshan Jiang, and Bingsheng He. Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024

    2024

  7. [15]

    From data to functa: Your data point is a function and you can treat it like one

    Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv:2201.12204, 2022

    2022 arXiv

  8. [16]

    Concentrated differential privacy

    Cynthia Dwork and Guy N Rothblum. Concentrated differential privacy. arXiv:1603.01887, 2016

    2016 arXiv

  9. [17]

    Our data, ourselves: Privacy via distributed noise generation

    Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006

    2006

  10. [18]

    On the convergence theory of gradient-based model-agnostic meta-learning algorithms

    Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. On the convergence theory of gradient-based model-agnostic meta-learning algorithms. In AISTATS, 2020 a

    2020

  11. [19]

    Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach

    Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. NeurIPS, 2020 b

    2020

  12. [20]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017

    2017

  13. [21]

    Property inference attacks on fully connected neural networks using permutation invariant representations

    Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM CCS, pages 619--633, 2018

    2018

  14. [22]

    Privacy-preserving collaborative learning with automatic transformation search

    Wei Gao, Shangwei Guo, Tianwei Zhang, Han Qiu, Yonggang Wen, and Yang Liu. Privacy-preserving collaborative learning with automatic transformation search. In CVPR, 2021

    2021

  15. [23]

    Generalizable neural fields as partially observed neural processes

    Jeffrey Gu, Kuan-Chieh Wang, and Serena Yeung. Generalizable neural fields as partially observed neural processes. In ICCV, 2023

    2023

  16. [24]

    Bounding training data reconstruction in dp-sgd

    Jamie Hayes, Borja Balle, and Saeed Mahloujifar. Bounding training data reconstruction in dp-sgd. In NeurIPS, 2023

    2023

  17. [25]

    Federated neural radiance fields

    Lachlan Holden, Feras Dayoub, David Harvey, and Tat-Jun Chin. Federated neural radiance fields. arXiv:2305.01163, 2023

    2023 arXiv

  18. [26]

    Fedavp: Augment local data via shared policy in federated learning

    Minui Hong, Junhyeog Yun, Insu Jeon, and Gunhee Kim. Fedavp: Augment local data via shared policy in federated learning. NeurIPS, 2024

    2024

  19. [27]

    Provable generalization of overparameterized meta-learning trained with sgd

    Yu Huang, Yingbin Liang, and Longbo Huang. Provable generalization of overparameterized meta-learning trained with sgd. NeurIPS, 2022

    2022

  20. [28]

    A survey on federated learning for resource-constrained iot devices

    Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained iot devices. IEEE Internet of Things Journal, 2021

    2021

  21. [29]

    Fedexp: Speeding up federated averaging via extrapolation

    Divyansh Jhunjhunwala, Shiqiang Wang, and Gauri Joshi. Fedexp: Speeding up federated averaging via extrapolation. arXiv:2301.09604, 2023

    2023 arXiv

  22. [30]

    Improving federated learning personalization via model agnostic meta learning

    Yihan Jiang, Jakub Kone c n \`y , Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv:1909.12488, 2019

    1909 arXiv

  23. [31]

    Scaffold: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, 2020

    2020

  24. [32]

    Generalizable implicit neural representations via instance pattern composers

    Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023

    2023

  25. [33]

    Communication-efficient federated learning with accelerated client gradient

    Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024

    2024

  26. [34]

    Scalable neural video representations with learnable positional features

    Subin Kim, Sihyun Yu, Jaeho Lee, and Jinwoo Shin. Scalable neural video representations with learnable positional features. NeurIPS, 2022

    2022

  27. [35]

    Locality-aware generalizable implicit neural representation

    Doyup Lee, Chiheon Kim, Minsu Cho, and WOOK SHIN HAN. Locality-aware generalizable implicit neural representation. NeurIPS, 2024

    2024

  28. [36]

    Differentially private meta-learning

    Jeffrey Li, Mikhail Khodak, Sebastian Caldas, and Ameet Talwalkar. Differentially private meta-learning. In ICLR, 2020

    2020

  29. [37]

    Federated learning on non-iid data silos: An experimental study

    Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In IEEE International Conference on Data Engineering, 2022

    2022

  30. [38]

    Meta-learning with adjoint methods

    Shibo Li, Zheng Wang, Akil Narayan, Robert Kirby, and Shandian Zhe. Meta-learning with adjoint methods. In AISTATS, pages 7239--7251. PMLR, 2023

    2023

  31. [39]

    On the convergence of fedavg on non-iid data

    Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv:1907.02189, 2019

    1907 arXiv

  32. [40]

    Meta-sgd: Learning to learn quickly for few-shot learning

    Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv:1707.09835, 2017

    2017 arXiv

  33. [41]

    Av-nerf: Learning neural fields for real-world audio-visual scene synthesis

    Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. NeurIPS, 2023

    2023

  34. [42]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In AISTATS, 2017

    2017

  35. [43]

    Golfdb: A video database for golf swing sequencing

    William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty, John McPhee, and Alexander Wong. Golfdb: A video database for golf swing sequencing. In CVPRW, 2019

    2019

  36. [44]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021

    2021

  37. [45]

    On first-order meta-learning algorithms

    Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv:1803.02999, 2018

    2018 arXiv

  38. [46]

    Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020

    2020

  39. [47]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019

    2019

  40. [48]

    Miner: Multiscale implicit neural representation

    Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In ECCV, 2022

    2022

  41. [49]

    Inr-v: A continuous representation space for video-based generative tasks

    Bipasha Sen, Aditya Agarwal, Vinay P Namboodiri, and CV Jawahar. Inr-v: A continuous representation space for video-based generative tasks. arXiv:2210.16579, 2022

    2022 arXiv

  42. [50]

    Petface: A large-scale dataset and benchmark for animal identification

    Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In ECCV, 2025

    2025

  43. [51]

    Membership inference attacks against machine learning models

    Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE Symposium on Security and Privacy, pages 3--18. IEEE, 2017

    2017

  44. [52]

    Implicit neural representations with periodic activation functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020

    2020

  45. [53]

    Adversarial generation of continuous images

    Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In CVPR, 2021

    2021

  46. [54]

    Inras: Implicit neural representation for audio scenes

    Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. NeurIPS, 2022

    2022

  47. [55]

    Federated learning for large-scale scene modeling with neural radiance fields

    Teppei Suzuki. Federated learning for large-scale scene modeling with neural radiance fields. arXiv:2309.06030, 2023

    2023 arXiv

  48. [56]

    Fed3dgs: Scalable 3d gaussian splatting with federated learning

    Teppei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv:2403.11460, 2024

    2024 arXiv

  49. [57]

    Learning large-scale neural fields via context pruned meta-learning

    Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, and Jonathan Richard Schwarz. Learning large-scale neural fields via context pruned meta-learning. NeurIPS, 2024

    2024

  50. [58]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020

    2020

  51. [59]

    Learned initializations for optimizing coordinate-based neural representations

    Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021

    2021

  52. [60]

    Decentnerfs: Decentralized neural radiance fields from crowdsourced images

    Zaid Tasneem, Akshat Dave, Abhishek Singh, Kushagra Tiwary, Praneeth Vepakomma, Ashok Veeraraghavan, and Ramesh Raskar. Decentnerfs: Decentralized neural radiance fields from crowdsourced images. arXiv:2403.13199, 2024

    2024 arXiv

  53. [61]

    State of the art on neural rendering

    Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo Martin-Brualla, Tomas Simon, Jason Saragih, Matthias Nie ner, et al. State of the art on neural rendering. In Computer Graphics Forum, 2020

    2020

  54. [62]

    Tackling the objective inconsistency problem in heterogeneous federated optimization

    Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. NeurIPS, 2020

    2020

  55. [63]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004

    2004

  56. [64]

    Yang, Farhad Farokhi, Shi Jin, Tony Q

    Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 2020

    2020

  57. [65]

    Federatedscope: A flexible federated learning platform for heterogeneity

    Yuexiang Xie, Zhen Wang, Dawei Gao, Daoyuan Chen, Liuyi Yao, Weirui Kuang, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope: A flexible federated learning platform for heterogeneity. arXiv:2204.05011, 2022

    2022 arXiv

  58. [66]

    Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction

    Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In CVPR, 2020

    2020

  59. [67]

    Gradient obfuscation gives a false sense of security in federated learning

    Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, and Huaiyu Dai. Gradient obfuscation gives a false sense of security in federated learning. In USENIX Security Symposium, 2023

    2023

  60. [68]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018

    2018

  61. [69]

    Federated neural radiance field for distributed intelligence

    Yintian Zhang and Ziyu Shao. Federated neural radiance field for distributed intelligence. arXiv:2406.10474, 2024

    2024 arXiv

  62. [70]

    Task-level differentially private meta learning

    Xinyu Zhou and Raef Bassily. Task-level differentially private meta learning. NeurIPS, 2022

    2022

  63. [71]

    Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction

    Hao Zhu, Haotian Yang, Longwei Guo, Yidi Zhang, Yanru Wang, Mingkai Huang, Menghua Wu, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction. IEEE TPAMI, 2023

    2023

  64. [72]

    Mofanerf: Morphable facial neural radiance field

    Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.