Pith. sign in

REVIEW 3 major objections 5 minor 17 references

Slice-level partial alignment of latent features, guided by partial information decomposition, lets heterogeneous agents in decentralized federated learning share knowledge without a server and without gradient surgery.

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

2026-08-03 10:25 UTC pith:BS7PZJFK

load-bearing objection The empirical work is genuinely broad, but the algorithm's own update rule contradicts its central 'partial alignment' claim: Eq. (8) exchanges the whole encoder and unique head, not just the redundant slice. the 3 major comments →

arxiv 2601.10012 v2 pith:BS7PZJFK submitted 2026-01-15 cs.LG

PID-Guided Partial Alignment for Multimodal Decentralized Federated Learning

Yanhang Shi , Xiaoyu Wang , Houwei Cao , Jian Li , Yong Liu This is my paper
classification cs.LG
keywords decentralized federated learningmultimodal learningpartial information decompositionfeature fissionpartial alignmentgradient conflictpeer-to-peermodality heterogeneity
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.

PARSE claims that the main obstacle in multimodal decentralized federated learning is not the absence of a server but the attempt to share a single monolithic representation across agents with different modality sets, which makes uni- and multimodal agents pull shared parameters in incompatible directions. The fix is to split each modality's latent into three slices — redundant, unique, and synergistic — following partial information decomposition, and to have agents exchange only the slices that are semantically alignable with their neighbors, over per-modality peer-to-peer subgraphs. This 'partial alignment' removes the need for central coordination or gradient-surgery techniques. The paper reports consistent accuracy gains over task-, modality-, and hybrid-sharing baselines across four benchmarks and diverse agent mixes and topologies, with the largest gains when multimodal agents are scarce. If correct, it shows that a simple server-free recipe — share only what is alignable, keep the rest local — is sufficient for heterogeneous-modality decentralized learning.

Core claim

The central claim is that the uni–/multimodal gradient conflict in decentralized federated learning can be resolved by feature fission plus partial alignment. Each agent's encoder latent is factorized into redundant, unique, and synergistic slices, guided by the PID identity I(X1,X2;Y)=U1+U2+R+S. Agents then communicate over per-modality subgraphs and exchange only the slices that are alignable given the modalities they share: unique and redundant slices among agents owning that modality, synergistic slices only among agents with the same full modality set. A contrastive-diversity loss aligns redundant slices across modalities of the same sample while treating unique/synergistic slices as ha

What carries the argument

Feature fission: each modality's encoder output is split into three equal slices zr, zu, zs (Eq. 3), intended to realize the PID decomposition of the information that the modalities carry about the label. The contrastive-diversity loss (Eq. 5) pulls redundant slices from different modalities of the same sample together, while the sample's own unique and synergistic slices serve as hard negatives, pushing the slices to be orthogonal. Communication runs on per-modality subgraphs with decentralized SGD: agents mix only the parameters for modalities they own (encoder, unique head, redundant head), and the synergistic classifier is exchanged only among agents with identical modality sets. This ro

Load-bearing premise

The equal three-way split of each encoder's latent, regularized by the contrastive-diversity loss, actually yields slices that carry redundant, unique, and synergistic information in the PID sense — yet no PID quantity is computed, and the split is a default choice rather than a derived one.

What would settle it

Compute an empirical PID estimate (e.g., the I_min/I_max formulas in the paper's appendix) on the learned zr, zu, zs slices. If the redundant slice does not capture the minimum per-modality label information and the synergistic slice does not exceed the best single-modality information, the slices do not realize the decomposition and the alignment story loses its grounding. A simpler behavioral test: shuffle the redundant slice across agents during training; if accuracy does not drop, the slice is not doing the alignment work.

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

If this is right

  • If PARSE is right, heterogeneous agents with different sensor/modality sets can collaborate effectively in a purely peer-to-peer network, so decentralized deployments (edge, IoT, autonomous systems) do not need a central aggregator for multimodal learning.
  • The method's gains grow as multimodal agents become scarce, so it is most valuable in realistic settings where most nodes have only one or a few modalities.
  • Per-link communication stays bounded because agents exchange only modality-specific parameters (encoder, unique/redundant heads); the synergistic head is a single linear layer shared only within same-modality-set groups.
  • The framework is topology-agnostic: it works on ring, chordal-ring, and random-gossip overlays, so no special network structure is required.
  • PARSE also transfers to server-based federated learning, where it is competitive with or better than dedicated server-coordinated multimodal FL methods, suggesting the slice design helps even when a server exists.

Where Pith is reading between the lines

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

  • The paper never computes a PID quantity, so the causal story depends on the slices actually being redundant/unique/synergistic; a direct information-theoretic check of the learned slices would either confirm the mechanism or reveal that the gains come from ensemble or regularization effects.
  • The equal-split default is explicitly a simple choice; a data-driven allocation of dimensions to slices (based on estimated R/U/S strengths) is a natural next step that could improve results further.
  • The principle of partial alignment — decompose representations into shareable and exclusive parts and communicate only the shareable part — could generalize beyond modality heterogeneity, e.g., to agents with different label spaces, tasks, or data distributions, wherever gradient conflict arises from heterogeneous information.
  • Because the contrastive-diversity loss is the only mechanism enforcing slice separation, a testable prediction is that removing the hard-negative structure would hurt unimodal agents and blur the synergy benefit; an ablation that treats unique/synergistic as positives should confirm this.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes PARSE, a server-free decentralized federated learning framework for multimodal agents with heterogeneous modality sets. PARSE splits each modality encoder's latent into redundant, unique, and synergistic slices (Eq. 3), applies a contrastive-diversity loss (Eq. 5), and communicates over per-modality subgraphs via DSGD (Eq. 8). The stated design is that only the redundant slice is shared across agents, while unique and synergistic components remain local. Experiments on KU-HAR, ModelNet-40, AVE, and IEMOCAP under several agent mixes, topologies, and non-IID levels, plus ablations on split ratios and fusion operators, report consistent accuracy gains over task-, modality-, and hybrid-sharing baselines.

Significance. If the slice-level partial alignment mechanism were realized as stated, this would be a substantial contribution: a server-free, topology-agnostic, gradient-surgery-free multimodal DFL method with a principled PID interpretation. The experimental scope is genuinely broad, including multiple benchmarks, baselines, agent ratios, topologies, and non-IID settings, and the paper also provides an FL comparison and ablation studies. However, the actual update rule synchronizes the full modality encoder and the unique head, so the central mechanism is not implemented as claimed. The PID grounding is asserted rather than verified, and several headline accuracy margins are inside one standard deviation. The contribution is therefore conditional on correcting the algorithm/interpretation gap and providing stronger statistical support.

major comments (3)
  1. [§3.4, Eq. (8)] The update rule mixes θ^m_i = {θ^m_{i,h}, θ^{m,u}_{i,f}, θ^r_{i,f}} across modality-m neighbors. Since h^m produces all three slices in Eq. (3), the entire encoder (and the unique head) is synchronized, not just the redundant slice/head. This directly contradicts §3.3, which states that 'only the redundant slice zr ... participates in cross-agent synchronization; zu and zs ... are optimized locally,' and it contradicts the abstract's 'partial alignment.' Additionally, §3.4's 'Synergistic-head Subgraphs' exchanges θ^s among agents with the same modality set, again not keeping synergy local. As implemented, PARSE is a shared-encoder method with auxiliary heads and losses, not a slice-level partial-alignment method. This is load-bearing for the paper's central claim and must be resolved by either changing the algorithm to truly share only redundant parameters or substantially reframing the
  2. [§2.2/§3.2 and Appendix I] The paper never computes a PID quantity. The equal split in Eq. (3) is a default choice, and the contrastive loss in Eq. (5) imposes orthogonality by construction; it does not establish that zr, zu, and zs carry redundant, unique, and synergistic information in the PID sense. The n-modality identities in Appendix I are not standard: Eq. (11) sums p(y) times pointwise minima, which is not the I_min redundancy of Williams & Beer, and Eq. (14) does not guarantee nonnegative unique information. Without a direct quantitative check, 'PID-guided' remains an interpretation rather than a derivation, and alternative explanations such as ensemble or regularization effects are not ruled out.
  3. [Table 2] Several headline margins are within one standard deviation of the strongest baseline, e.g., KU-HAR [A] 80.6±1.0 vs 80.1±0.7, AVE [G] 53.3±1.5 vs 52.4±0.2, and IEMOCAP [A] 48.2±0.6 vs 47.7±0.6. Since the paper claims PARSE 'consistently outperforms' all baselines, it should report paired significance tests or repeated-seed distributions for the main α=0.5 ring results. This is a rigor issue, not by itself disqualifying, but it affects the strength of the empirical claim.
minor comments (5)
  1. [§4, comparison methods paragraph] Typo: 'omstamtoated' should be 'instantiated'.
  2. [Appendix B] The sentence 'The batch size is set to 32 across all experiments.' appears twice in consecutive lines.
  3. [§3.2] 'Motivate by PID' should be 'Motivated by PID'.
  4. [Table 5] The non-IID comparison reports mean accuracies without standard deviations or seed counts; please add error bars or state the number of runs.
  5. [§3.4, Synergistic-head Subgraphs] The statement that synergistic parameters are exchanged only among agents sharing the same modality set should be reconciled with the paper's claim that synergy is learned locally; this is related to major comment 1.

Circularity Check

0 steps flagged

No significant circularity: the empirical comparisons are self-contained; the PID slice labels are asserted rather than derived, which is a grounding weakness but not a circular reduction.

full rationale

The paper's central empirical claim—consistent accuracy gains over six DFL baselines across four benchmarks—rests on external, self-contained comparisons. No baseline number is a fitted version of PARSE's own output, and no reported gain is computed from a parameter fit to that gain. The PID framing is motivational: Eq. (3) splits a latent by concatenation, Eq. (5) aligns the first slice and repels the other two, and the paper then names the slices redundant/unique/synergistic. This naming is definitional rather than a derivation from Eq. (1), and no PID quantity is actually computed; the mechanism story is therefore under-supported as a matter of construct validity, but it is not a circular reduction of a prediction to an input. The paper also explicitly treats the split as a default ('We adopt an equal split by default', §3.2) and concedes that data-driven allocation is future work (Appendix E), which confirms the split is a design choice rather than a fitted quantity disguised as a result. There is no load-bearing self-citation chain: the PID references are external, and no uniqueness theorem from the authors' prior work is invoked. One internal inconsistency—§3.3 says only the redundant slice participates in synchronization, while Eq. (8) mixes the full modality encoder plus the unique head—is a real implementation/claim mismatch and a correctness risk, but it is not an equivalence between a prediction and an input, so it does not count as circularity under the specified criteria.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The paper contributes a routing mechanism (slice-level partial alignment) and two design devices (slices, virtual agents), but pulls the PID identity and pretrained encoders from the literature. The honest ledger: 3 hyperparameters tuned on the benchmarks, a hand-chosen split ratio, and 5 axioms — the most fragile being that an arbitrary equal split plus a contrastive loss realizes PID.

free parameters (4)
  • β (contrastive-diversity loss weight) = 0.2
    β is swept on KU-HAR (Table 6, range 0.0–0.4) and the argmax 0.2 is used across all datasets; the paper notes performance declines for β>0.2, so the reported gains partly depend on this tuned value.
  • τ (contrastive temperature) = 0.2
    Appendix C.4 sets τ=0.2 'throughout all experiments'; no sensitivity analysis is reported, so it is a hand-chosen constant the central contrastive objective depends on.
  • Latent split ratio (dims of zr, zu, zs) = Equal thirds (64/64/64 of 192 on KU-HAR; 128/128/128 of 384 on AVE and IEMOCAP; 100/100/100 on ModelNet-40)
    The default even split is justified via Appendix E sweeps, but those sweeps show dataset-dependent optima (KU-HAR synergy peaks at 96d with combined 89.5), so the default is a hand-selected compromise, not derived from PID.
  • Per-dataset learning rates = 0.1 (KU-HAR), 0.01 (ModelNet-40), 0.005 (AVE), 5e-5 Adam (IEMOCAP)
    Standard per-dataset tuning (Appendix C.4); applied equally to baselines, so it does not bias relative comparisons, but it is a free choice affecting all absolute numbers.
axioms (5)
  • ad hoc to paper The equal-split latent slices zr/zu/zs trained with Eq. (5) realize the PID decomposition of Eq. (1).
    Postulated in §3.2–3.3; no mutual information is ever estimated, so the mapping from 'redundant/unique/synergistic' to actual information terms is asserted.
  • domain assumption Gradient cosine misalignment (Fig. 3) is the causal factor behind hybrid-sharing failures, and slice isolation is the reason PARSE fixes them.
    Section 2.1 establishes correlation between inter-group gradient angles and performance; no intervention experiment rules out confounds such as differing local loss objectives.
  • ad hoc to paper Exchanging only the redundant slice and head f_r across agents is sufficient for stable consensus; unique and synergistic branches can be trained purely locally.
    Core of partial alignment (§3.3–3.4); no convergence or bias analysis is provided for the partial averaging update in Eq. (8).
  • ad hoc to paper The n-modality PID identities in Appendix I (Eqs. 10–14), based on pointwise I_min/I_max, are valid definitions of redundancy/synergy/unique information.
    These formulas are approximations presented as derivations; the standard PID framework (Griffith & Koch 2014) requires an optimization over distributions, which is not carried out, and the main text never uses them.
  • domain assumption Pretrained backbones (MobileNetV3-Large/Small, BERT-large, DenseNet) provide modality features that transfer to the decentralized non-IID regime.
    Used in Appendix C for AVE/ModelNet-40/IEMOCAP; all baselines share them, so relative comparisons are fair, but absolute numbers depend on this transfer.
invented entities (2)
  • Redundant/unique/synergistic latent slices (zr, zu, zs) no independent evidence
    purpose: Routing primitive: determines what is shared (zr) vs. kept local (zu, zs) across peers.
    Internal representations of the proposed model; the only 'evidence' is t-SNE of the model's own outputs (Fig. 8), so there is no falsifiable handle outside the paper.
  • Virtual unimodal agents no independent evidence
    purpose: Represent each multimodal agent as one virtual agent per owned modality so it can join per-modality subgraphs (§3.4).
    Bookkeeping construct; consistent with modality-based sharing but with no independent validation that it preserves convergence or consensus properties.

pith-pipeline@v1.3.0-alltime-deepseek · 32955 in / 16896 out tokens · 158595 ms · 2026-08-03T10:25:41.067542+00:00 · methodology

0 comments
read the original abstract

Multimodal decentralized federated learning (DFL) must support collaboration among agents that hold different modality subsets and often different model components, while operating over peer-to-peer (P2P) overlays without a coordinating server or a global network view. A key obstacle is that conventional multimodal training often relies on a single shared representation, which implicitly assumes that heterogeneous peers can exchange and aggregate the same model components over the same communication links. In multimodal DFL, this assumption breaks down: uni- and multimodal agents may push incompatible updates through shared overlays, weakening both inter-agent transfer and cross-modal interaction. We present PARSE, a server-free framework that brings partial information decomposition (PID) into multimodal DFL. Each agent splits its latent features into redundant, unique, and synergistic slices ("feature fission"), and performs slice-aware communication over modality-conditioned P2P overlays. During training, agents exchange only the slices that are semantically alignable with their neighbors, according to the modalities and model components they share ("partial alignment"). This design avoids centralized orchestration and gradient-surgery style conflict handling, while remaining compatible with standard DFL constraints and a range of P2P overlay topologies. Across multiple benchmarks and heterogeneous peer mixes, PARSE consistently outperforms task-, modality-, and hybrid-sharing multimodal DFL baselines while keeping per-link payloads bounded. Ablations on fusion choices and split ratios, together with qualitative feature analyses and overlay-topology studies, demonstrate the robustness and communication efficiency of the proposed slice-aware design.

Figures

Figures reproduced from arXiv: 2601.10012 by Houwei Cao, Jian Li, Xiaoyu Wang, Yanhang Shi, Yong Liu.

Figure 2
Figure 2. Figure 2: Test accuracy for unimodal and multimodal agents under task-, modality-, and hybrid [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: We study three knowledge-sharing strate [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Group agents by their modality subsets, we show inter-group gradient co￾sine similarity (w.r.t. shared models) and intra-group gradi￾ent cosine similarity. • Hybrid sharing fails to fully exploit multimodal interaction. It achieves only marginal gains for unimodal agents over task-based sharing, with modest multimodal improvements as the unimodal ratio increases ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: PARSE at a glance (two modalities for illustration). Agents that share a modality form a modality-specific P2P subgraph. Each encoder output is fissioned into redundant (z r ), synergistic (z s ), and unique (z u ) slices. Unimodal agents train on z u and z r . Multimodal agents: (i) partially align z u and z r across agents, and (ii) learn a multimodal classifier on the fused z s . This routes shareable i… view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE visualization of synergistic features from each modality and after fusion. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of different methods under various communication topologies. Impact of Overlay Topology. We evaluate PARSE and all baselines on three P2P overlay topologies: Ring (Koloskova et al., 2019); Chordal Ring (Ahmad et al., 2022), which augments each node with a dia￾metrically opposite link and roughly halves the diam￾eter; and Random Gossip (De Vos et al., 2023), where each round each agent exchanges … view at source ↗
Figure 7
Figure 7. Figure 7: Impact of ensemble pre￾dictions in the classification loss. Ensemble Training Analysis. We compare three schemes: Ensemble-only (ours): a single loss on the sum of all three component predictions; Component-only: independent losses for each component; and Ensemble+Component: the ensemble loss plus three per-component losses [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: t-SNE visualization of synergistic features from each modality and after fusion, showing [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 4 linked inside Pith

  1. [1]

    Studies of chordal ring networks via double metric dimensions.Mathematical Problems in Engineering, 2022(1):8303242,

    M Ahmad, Z Zahid, M Javaid, and E Bonyah. Studies of chordal ring networks via double metric dimensions.Mathematical Problems in Engineering, 2022(1):8303242,

    2022

  2. [2]

    Each ring is formed either (i) by agents that share a given modality (modality-specific ring) or (ii) by agents that have the exact same modality set (task-specific ring)

    Unless stated otherwise, agents are linked in a ring topology (Koloskova et al., 2019). Each ring is formed either (i) by agents that share a given modality (modality-specific ring) or (ii) by agents that have the exact same modality set (task-specific ring). During every communication round, each agent runs one local training epoch and then exchanges its...

    2019

  3. [4]

    (b) KU-HAR gyroscope

    Figure 8 presents t-SNE plots of the synergistic features 21 Preprint (a) KU-HAR accelerometer. (b) KU-HAR gyroscope. (c) KU-HAR fused features. (d) A VE audio. (e) A VE video. (f) A VE fused features. (g) ModelNet-40 V1. (h) ModelNet-40 V2. (i) ModelNet-40 fused features. (j) IEMOCAP audio. (k) IEMOCAP video. (l) IEMOCAP text. (m) IEMOCAP fused fea- ture...

    2022

  4. [6]

    Nonnegative decomposition of multivariate information.arXiv preprint arXiv:1004.2515,

    Paul L Williams and Randall D Beer. Nonnegative decomposition of multivariate information.arXiv preprint arXiv:1004.2515,

    Pith/arXiv arXiv

  5. [7]

    Communication-efficient multimodal federated learning: Joint modality and client selection

    Liangqi Yuan, Dong-Jun Han, Su Wang, Devesh Upadhyay, and Christopher G Brinton. Communication-efficient multimodal federated learning: Joint modality and client selection. arXiv preprint arXiv:2401.16685, 2024a. Liangqi Yuan, Ziran Wang, Lichao Sun, Philip S Yu, and Christopher G Brinton. Decentralized federated learning: A survey and perspective.IEEE In...

    arXiv

  6. [8]

    A RELATEDWORK We review three closely related research directions: multimodal representation learning, multimodal federated learning, and the emerging field of multimodaldecentralizedfederated learning. Multimodal Representation Learning.Early work merges modalities via simple concatenation or cross-attention, yielding a single embedding that blends redun...

    2020

  7. [9]

    But all assume centralized, full-modality access and a monolithic backbone—impractical under privacy or bandwidth constraints

    and ALIGN (Jia et al., 2021), or via information- theoretic decomposition (Liang et al., 2023). But all assume centralized, full-modality access and a monolithic backbone—impractical under privacy or bandwidth constraints. In contrast, our feature- fission perspective is orthogonal and complementary: rather than learning a single embedding, we separate ea...

    2021

  8. [10]

    It aligns modality-shared knowledge across agents by broadcasting a feature generator over afully-connectedpeer-to-peer network

    is, to the best of our knowledge, the first algorithm expressly designed for multimodal DFL. It aligns modality-shared knowledge across agents by broadcasting a feature generator over afully-connectedpeer-to-peer network. However, it does not explicitly account for synergistic interactions and heavily relies on the presence of a sufficient number of multi...

    2018

  9. [16]

    As shown, even without a dedicated server-side design,PARSE achieves leading performance across all multimodal agents (notably on IEMOCAP) and most unimodal agents (with the exception of text-only agents on IEMOCAP). These findings highlight the strong extension potential of PARSE in standard FL settings, suggesting that further gains may be realized by c...

    2014

  10. [64]

    Hardware.All experiments are conducted on a machine equipped with a Tesla V100-SXM2 GPU (32GB VRAM) and a Intel Xeon Silver 4216 CPU (2.1 GHz, 16 cores, 22MB cache)

    By default, we split each dataset into 80% for training and 20% for testing. Hardware.All experiments are conducted on a machine equipped with a Tesla V100-SXM2 GPU (32GB VRAM) and a Intel Xeon Silver 4216 CPU (2.1 GHz, 16 cores, 22MB cache). Methods.For PARSE, we decompose the input features into three components. The feature dimensions are split from 38...

    2021

  11. [353]

    IEMOCAP.We use the same preprocessing procedure as in Liang et al

    The resulting spectrogram is normalized per segment using zero mean and unit variance. IEMOCAP.We use the same preprocessing procedure as in Liang et al. (2020). We use three modalities: acoustic, visual, and textual.Acoustic featuresare extracted using the openSMILE toolkit, producing 130-dimensional frame-level descriptors. To reduce redundancy, we down...

    2020

  12. [384]

    Table 14: Performance (Acc.) vs

    All other settings follow our default ablation protocol. Table 14: Performance (Acc.) vs. feature split size across datasets. (a) KU-HAR Varying Split Metric 32d 64d 96d 128d Unique Unique-only 83.5 86.2 86.7 86.1 Combined 87.8 88.6 88.8 88.5 Redundant Redundant-only 87.2 87.8 87.4 87.6 Combined 88.7 88.6 87.6 87.0 Synergistic Synergistic-only 54.5 61.5 6...

    2023

  13. [2008]

    Following Liang et al

    is an emotion-recognition corpus with audio, visual, and text modalities A, V , T. Following Liang et al. (2020), we keep four emotion labels—happy,sad,angry, andneutral. Communication setup.We follow the modality- and task-based connection schemes described in Section

    2020

  14. [2014]

    M3h: Multimodal multitask machine learning for healthcare.arXiv preprint arXiv:2404.18975,

    Dimitris Bertsimas and Yu Ma. M3h: Multimodal multitask machine learning for healthcare.arXiv preprint arXiv:2404.18975,

    Pith/arXiv arXiv

  15. [2019]

    Measuring the effects of non-identical data distribution for federated visual classification.arXiv preprint arXiv:1909.06335,

    Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification.arXiv preprint arXiv:1909.06335,

    Pith/arXiv arXiv 1909

  16. [2021]

    Fedclip: Fast generalization and personalization for clip in federated learning.arXiv preprint arXiv:2302.13485,

    Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning.arXiv preprint arXiv:2302.13485,

    Pith/arXiv arXiv

  17. [2024]

    Towards optimal multi-modal federated learning on non-iid data with hierarchical gradient blending

    Sijia Chen and Baochun Li. Towards optimal multi-modal federated learning on non-iid data with hierarchical gradient blending. InIEEE INFOCOM 2022-IEEE conference on computer communications,

    2022