acceptodds
Under review as a conference paper at ICLR 2027

FedAffine: Personalized Federated Learning with Shared Affine Predictor Families

Abstract

Personalized federated learning must decide which cross-client differences should survive collaboration, yet trained client models conflate persistent population variation with finite-sample estimation error. We call this the population–estimation ambiguity and impose a structural restriction: client population predictors lie near a low-dimensional affine family. A sample-weighted sequence model decomposes risk into unrepresented population variation, client-specific estimation, and pooled estimation, separating personalization capacity from coefficient regularization. This motivates FedAffine, which collaboratively learns a high-dimensional base and shared basis while keeping a short coefficient local. Coefficients can be reconstructed from local data, enabling deployment to held-out clients without persistent training-time state. Under the pFL-Bench FEMNIST protocol, reconstructive FedAffine re-estimates three coefficients from zero and attains weighted accuracy of on participating clients and on held-out clients, versus and for Ditto. We additionally report the accuracy–adaptation-compute frontier. On multi-institutional radiotherapy dose prediction, FedAffine achieves site-macro Metric 1 of Gy, compared with for FedAvg-FT. Controlled experiments validate the fixed-family risk calculation and find predictive but incomplete alignment with a planted span; the analysis does not guarantee neural recovery of that span.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
personalized federated learning, client heterogeneity, affine predictor families

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_EeDCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedAffine, a personalized federated learning framework based on shared affine predictor families. It learns a shared base and basis functions while maintaining low-dimensional client-specific coefficients for personalization. A statistical analysis characterizes the trade-off between population variation and estimation error. Experiments on image classification and radiotherapy dose prediction evaluate its effectiveness across different federated settings.

Critical strengths
  1. The paper offers an interesting perspective on personalized federated learning by explicitly distinguishing persistent population differences from finite-sample estimation errors.

  2. The method demonstrates promising results on FEMNIST and real-world radiotherapy dose prediction, with effective personalization for held-out clients.

Critical weaknesses
  1. The methodological novelty is limited, as the shared affine predictor structure overlaps with existing approaches. Its technical distinction remains insufficiently justified.

  2. The theoretical analysis does not directly support the neural implementation, while the inconsistent improvements across datasets weaken the empirical contribution.

Questions for authors
  1. What distinguishes FedAffine from existing linear hypernetwork and output-adaptation methods?

  2. What evidence supports that the learned basis captures persistent population differences rather than estimation noise?

  3. Why does FedAffine underperform Ditto and FedAvg-FT on CIFAR-10/100?

  4. How do the different coefficient-handling protocols affect the generality of the proposed framework?

Reviewer_sMWWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedAffine, a personalized federated learning framework that represents client-specific predictors as affine combinations of a shared base function and a small number of learned basis functions. The motivation is the population–estimation ambiguity: observed differences between client models may arise from genuine population heterogeneity or finite-sample estimation noise.

The method collaboratively optimizes the shared base and basis functions while estimating low-dimensional client-specific coefficients from local labeled data. The authors also derive an approximation–estimation risk decomposition under a fixed-family Gaussian sequence model, yielding an oracle criterion for selecting personalization directions.

Experiments are conducted on FEMNIST, Office-Caltech, CIFAR-10/100, and multi-institutional radiotherapy dose prediction. FedAffine achieves competitive results on FEMNIST and improves dose-prediction accuracy over several established federated learning baselines. However, it underperforms full-model personalization methods on CIFAR-10/100.

Critical strengths

S1. A useful and well-articulated statistical perspective on personalization.

The paper explicitly distinguishes persistent population heterogeneity from finite-sample estimation error. The fixed-family analysis decomposes prediction error into population approximation error, client-specific estimation error, and pooled estimation error. This provides an interpretable framework for understanding the trade-off between personalization capacity and statistical efficiency.

I particularly appreciate that the authors carefully state the assumptions and limitations of their theoretical results rather than presenting the oracle analysis as a guarantee for neural network training.

S2. Encouraging empirical results on realistic federated learning tasks.

The experimental evaluation goes beyond synthetic client partitions and standard image classification. In particular, the multi-institutional radiotherapy dose-prediction experiment provides a meaningful application involving naturally occurring client heterogeneity.

FedAffine reduces the site-macro dose-prediction error from 4.03 Gy for FedAvg-FT and 4.00 Gy for Ditto to 3.52 Gy. The FEMNIST results also demonstrate promising performance on held-out clients, with 90.22% weighted accuracy compared with 89.41% for Ditto.

The coefficient-shuffling, random-basis, and adaptation-compute analyses provide additional evidence that the learned shared basis and client-specific coefficients are useful for prediction.

Critical weaknesses

W1. Limited methodological novelty and an insufficient connection between the theoretical analysis and the learned neural predictor family.

The central parameterization is closely related to existing shared-basis multitask learning, low-rank personalized models, and shared-representation methods with constrained client-specific prediction heads.

The paper acknowledges connections to GO-MTL, ELLA, FedRep, and linear hypernetworks. However, it remains unclear whether the proposed approach introduces a sufficiently distinct methodological contribution beyond applying a shared low-dimensional output basis with locally optimized coefficients in federated learning.

Moreover, the main theoretical results concern a fixed predictor family with independent, isotropic estimation errors. The optimal population subspace is assumed available for the oracle analysis, whereas FedAffine must learn its basis jointly with the shared neural model from finite client datasets.

This distinction is important because the proposed framework is motivated by separating population variation from estimation noise, yet the actual algorithm does not explicitly estimate these two sources of variation or guarantee their separation.

The neural planted-span experiment further illustrates this limitation: even under exact specification, the reported normalized subspace overlap is only 0.29 ± 0.10 (Table 14). Although predictive performance improves, the results do not establish that the learned basis identifies persistent population heterogeneity.

Consequently, the theory provides useful intuition but does not yet explain why the proposed neural optimization procedure should reliably learn the intended population structure.

W2. Experimental comparisons do not fully isolate the advantage of the affine structure, and performance is inconsistent across heterogeneity settings.

Several experimental design choices make it difficult to attribute the reported gains specifically to the proposed affine predictor family.

First, the main FEMNIST comparison uses different personalization and adaptation procedures across methods. FedAffine performs 200 coefficient-estimation steps at evaluation, whereas the primary FedAvg-FT baseline follows a released five-step full-model adaptation procedure. Although the paper provides additional compute-frontier analyses, Appendix D.1.2 also shows that unrestricted post-hoc adaptation from a FedAvg checkpoint can achieve comparable accuracy. A more controlled comparison against equally tuned and compute-matched full-model and low-rank adaptation baselines is needed.

Second, the GDP-HMM implementation initializes the shared affine model from a pretrained FedAvg checkpoint before the reported federated training. It is unclear whether the computational cost of obtaining this initialization is included in the overall training budget. Furthermore, the medical experiment does not include a sufficiently direct parameter- and compute-matched control that isolates the learned affine basis from the additional residual prediction capacity and optimization procedure.

Third, FedAffine is not consistently competitive across datasets. On CIFAR-100, it achieves 60.96% weighted accuracy, substantially below Ditto (66.53%) and FedAvg-FT (66.60%). On Office-Caltech, the improvement over FedBN is relatively modest, and the authors report that local batch-normalization is important for the best operating point.

These results suggest that the usefulness of the affine restriction depends strongly on the form of client heterogeneity. However, the paper does not sufficiently characterize when the assumed low-dimensional population structure is appropriate or when more flexible personalization is preferable.

Questions for authors

Q1. What is the essential methodological distinction from existing shared-basis and low-rank personalization approaches? Q2. How sensitive are the results to the total training and adaptation budgets? Q3. Can the authors provide stronger evidence that the learned basis captures persistent population variation rather than finite-sample estimation artifacts?

Reviewer_b3cAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposed FedAffine which is a personalized federated learning method that represents client predictors using a shared base, a shared low-rank basis, and locally estimated coefficients. It basically derives an approximation estimation trade-off for a fixed affine family and evaluates the approach on federated image classification and multi-institutional radiotherapy dose prediction. While the formulation is intuitive and the FEMNIST results are partly promising, the theoretical guarantees do not directly extend to the learned neural implementation, and the empirical evidence does not consistently demonstrate an advantage over existing personalization methods. Also, the writing style of paper is not clear to understand.

Critical strengths

The proposed method has a clear mathematical formulation, separates representational capacity from local coefficient fitting, and provides a useful fixed family risk decomposition. The evaluation includes held-out clients and an adaptation compute comparison, and the paper reports limitations of the neural span recovery experiments. The GDP-HMM results also provide an interesting application beyond standard image classification.

Critical weaknesses

The primary concern that I have is the gap between the paper's statistical motivation and its empirical validation. The theory assumes a fixed family and isotropic estimation noise.However the proposed method jointly learns a neural basis under stochastic federated optimization. The paper does not sufficienctly establish that the learned basis captures persistent population variation rather than correlated estimation artifacts. Also, FedAffine [the proposed solution] underperforms strong baselines on CIFAR-10/100, and the plante span experiment demonstrates only partial subspace alignment. These issues limit the evidence for both the central mechanism and the method's general applicability.

Questions for authors

Can the authors quantify the stability of the learned basis across independent training samples from the same client populations? How does proposed method compare with a parameter-matched linear output adaptor using identical coefficient initialization and adaptation budgets? Can the authors please provide a practical, data-driven rank-selection procedure? How robust are the results when the client estimation errors are correlated or anisotropic? Do the conclusions hold under matched adaptation-label budgets and equivalent total compute?

Reviewer_yhEQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedAffine, a personalized federated learning method designed to improve prediction accuracy under client heterogeneity by modeling client predictors as members of a shared low-dimensional affine family. Instead of learning a single global model or fully independent personalized models, FedAffine collaboratively learns a shared base predictor and a set of basis functions, while each client estimates a small local coefficient vector to construct its personalized predictor. The method is evaluated on FEMNIST, Office-Caltech, label-skewed CIFAR-10/100, and multi-institutional 3D radiotherapy dose prediction. The result shows that FedAffine achieve better personalized prediction accuracy, including for held-out clients after local coefficient reconstruction.

Critical strengths

S1. The paper starts from a fundamental question in pFL, which differences between client models represent persistent population heterogeneity and should be preserved, and which differences are merely finite-sample estimation noise and should be suppressed through collaboration. The authors clearly define this using the term "population–estimation ambiguity". It opens a new directions for understanding the heterogeneity of data.

S2-W1. The paper evaluates beyond standard classification benchmarks. Experiments also consider several dimensions including: different datasets and estimate both seen and unseen clients. The authors also provide useful ablation study. However, this strong points is also one of the weak points where most of the results and discussion are placed in the supplemental material. The author should add more deep discussion into the main manuscript

Critical weaknesses

W1 (See S2-W1)

W2. As the author mentioned in the conclusion, "The broader image benchmarks show that the affine restriction can help under domain shift but need not outperform full-model personalization under severe label skew." Label skew is one of the most data skew frequent appear in practical. It narrow down the impact of the proposed method. Could the authors discuss more in this points. Is there any direction to fix this problem?

W3. Incomplete coverage of related work. The paper compares FedAffine with other methods for unseen clients but does not discuss the paper “Enhancing the Generalization of Personalized Federated Learning with Multi-head Model and Ensemble Voting.” The authors should expand the related-work discussion and provide an appropriate comparison.

Questions for authors

Q1. How should practitioners select the affine rank rr without access to held-out test clients?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.