Home / Curated / Algorithms / FedPAC (Personalized Federated Learning with Feature Alignment and Classifier Collaboration)

FedPAC (Personalized Federated Learning with Feature Alignment and Classifier Collaboration)

Algorithms AIMech:FedPacFeatureAlignmentAndClassifierCollaborationForPersonalizedFederatedLearning DRAFT

Aliases: FedPAC, Feature Alignment and Classifier Collaboration in Personalized Federated Learning, Quadratic Bias-Variance Classifier Collaboration for Personalized Federated Learning

A personalized federated learning algorithm (Xu, Tong, & Huang 2023, Sections 3-4, Theorem 1, Equations 3-4, 10-19, 43-46) that jointly (1) aligns local feature extractors f_{theta_i} to global class feature centroids c_k via the regularization penalty R_i(theta_i; c) = (lambda / n_i) sum_{l=1}^{n_i} (1 / d) ||f_{theta_i}(x_l) - c_{y_l}||_2^2 (Equations 3, 17, 19) to reduce the local-global feature generalization gap Delta R_i (Equations 12-14), and (2) constructs each client i's personalized classifier head phi_hat_i^{(t+1)} = sum_{j=1}^m alpha_{ij} phi_j^{(t+1)} (Equation 4) using simplex-constrained weights alpha_i^* that minimize the expected chi_i^2 testing loss bound via the convex quadratic program min_{alpha_i >= 0, sum_j alpha_{ij} = 1} sum_{j=1}^m (alpha_{ij}^2 / n_j) V_j + sum_{j, j'=1}^m alpha_{ij} alpha_{ij'} D_{jj'} (Theorem 1, Equations 10-11, 43-46) computed from first- and second-order client feature statistics.

Identity Notes & Concept Scope Distinct from FedRep (AIMech:FedRepSharedRepresentationPersonalizedFederatedLearning, Collins et al. 2021) and FedPer (AIMech:FedPerFederatedLearningWithPersonalizationLayers), which keep each client's classifier head strictly local without inter-client classifier collaboration or global centroid alignment; and distinct from FedProto (AIMech:FedProtoFederatedPrototypeLearning, Tan et al. 2022), which communicates only class prototypes without aggregating shared feature extractors or solving a bias-variance quadratic program for personalized classifier combinations.

Claims & Grounded Evidence

4 Claims

Mechanism #r460-quadratic-bias-variance-classifier-collaboration-qp-mechanism

Decomposes the expected chi_i^2 testing risk of a linearly combined classifier g_hat_i^* = sum_{j=1}^m alpha_{ij} g_hat_j into a pairwise distribution-mismatch bias quadratic form sum_{j,j'} alpha_{ij} alpha_{ij'} D_{jj'} and a sample-variance term sum_j (alpha_{ij}^2 / n_j) V_j (Theorem 1, Equations 10-11), and estimates V_j and D_{jj'} server-side from client class-conditional feature means mu_{j,y} and trace second moments tr(Lambda_{j,y}) under isotropic feature normalization (Appendix A.5, Equations 43-46) to solve a convex quadratic program for alpha_i^*.

supports Source: Personalized Federated Learning with Feature Alignment and Classifier Collaboration · Locator: Section 4.1 Theorem 1 (Equations 10-11) & Appendix A.2-A.5 (Equations 29-35, 43-46), pp. 5-6, 15-18

Scope #r460-global-centroid-feature-alignment-generalization-scope

Bounds the local generalization gap Delta R_i = 2 E_{P_Y^{(i)}}[(E_{P_{X|Y}^{(i)}}[f(x)] - E_{P_hat_{X|Y}^{(i)}}[f(x)])^T g_i(Y) / K] (Equations 12-14) by pulling local features f_{theta_i}(x_l) toward globally aggregated class centroids c_{y_l} (Equations 3, 17, 19), ensuring that classifier heads trained on different clients operate in a common aligned semantic subspace.

supports Source: Personalized Federated Learning with Feature Alignment and Classifier Collaboration · Locator: Section 3.2 (Equation 3) & Section 4.2 Feature Alignment by Regularization (Equations 12-19), pp. 4, 6-7

Limitation #r460-feature-statistics-upload-and-qp-overhead-limitation

Requires clients to compute and upload class-conditional feature means mu_{i,y} and scalar variance statistics V_i in addition to model parameters, and invokes a server-side m-variable quadratic program for each participating client during the initial collaboration rounds.

supports Source: Personalized Federated Learning with Feature Alignment and Classifier Collaboration · Locator: Section 4.2 Practical Implementation & Appendix A.5 Privacy and Communication Discussion (Equations 43-46), pp. 7, 17-18

Application #r460-personalized-vision-classification-under-label-skew-application

Applied to personalized federated image classification on EMNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and CINIC-10 under pathological and Dirichlet non-IID label distributions.

supports Source: Personalized Federated Learning with Feature Alignment and Classifier Collaboration · Locator: Section 5 Experiments (Tables 1-3, Figures 2-4), pp. 7-9

⚠️ Curated Limitations & Negative Results

Requires transmitting class-conditional feature centroids mu_{i,y} and scalar variance statistics V_i between clients and the server.
Relies on an isotropic feature-covariance approximation Lambda_f^{(i)} approx I_d to avoid O(d^2) covariance matrix transmission.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	narrow_match	FedPAC is a personalized federated learning algorithm combining global centroid feature alignment with QP-optimized inter-client classifier collaboration.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r460-primary-pdf`	Personalized Federated Learning with Feature Alignment and Classifier Collaboration Jian Xu, Xinyi Tong, Shao-Lun Huang	primary paper full_text	ID: https://arxiv.org/abs/2306.11867 sha256:88ac11e75187…
`r460-arxiv-abs`	Personalized Federated Learning with Feature Alignment and Classifier Collaboration (arXiv abstract) Jian Xu, Xinyi Tong, Shao-Lun Huang	primary paper abstract_only	ID: https://arxiv.org/abs/2306.11867 sha256:16a7433e22d2…
`r460-metadata-cr`	Personalized Federated Learning with Feature Alignment and Classifier Collaboration (Crossref metadata) Jian Xu, Xinyi Tong, Shao-Lun Huang	primary paper abstract_only	ID: https://arxiv.org/abs/2306.11867 sha256:97b8e3687eba…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:03:00Z

Updated: 2026-10-19T09:03:00Z

Source File: data/curated/FedPacFeatureAlignmentAndClassifierCollaborationForPersonalizedFederatedLearning.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories