Home / Curated / Algorithms / FedALA (Adaptive Local Aggregation for Personalized Federated Learning)

FedALA (Adaptive Local Aggregation for Personalized Federated Learning)

Algorithms AIMech:FedAlaAdaptiveLocalAggregationForPersonalizedFederatedLearning DRAFT

Aliases: FedALA, Federated Learning with Adaptive Local Aggregation, Adaptive Local Aggregation for Personalized Federated Learning

A personalized federated learning algorithm (Zhang et al. 2023, Section 3, Equations 1-6, Algorithm 1) whose Adaptive Local Aggregation (ALA) module initializes each client i's local model Theta^hat_i^t before local training in iteration t by copying the lower |Theta_i| - p generic feature-extractor layers directly from the server-broadcast global model Theta^{t-1} while element-wise interpolating the top p task-specific layers between the client's previous local model Theta_i^{t-1} and the global model Theta^{t-1} via learnable client-specific aggregation weights W_i^p clipped to [0, 1] by sigma(w) = max(0, min(1, w)): Theta^hat_i^t := Theta_i^{t-1} + (Theta^{t-1} - Theta_i^{t-1}) odot [1^{|Theta_i|-p}; sigma(W_i^p)] (Equation 4), where W_i^p is trained on an s% random subset D~_i^t of local data via W_i^p <- W_i^p - eta * (dL_i^t / dTheta^hat_i^{t,p}) odot (Theta^{t-1} - Theta_i^{t-1})^p with Theta^{t-1} and Theta_i^{t-1} frozen (Equation 5, Appendix A Equation 7).

Identity Notes & Concept Scope Distinct from APFL (Adaptive Personalized Federated Learning, Deng et al. 2020), which mixes a separately trained local model and global model using a single scalar mixing coefficient alpha_i in [0, 1] after or parallel to full-model training; and distinct from FedPer (AIMech:FedPerFederatedLearningWithPersonalizationLayers) and FedRep (AIMech:FedRepSharedRepresentationPersonalizedFederatedLearning, Collins et al. 2021), which hard-code binary layer splits (sigma in {0, 1}) and discard all global information in the personalized head layers rather than learning element-wise [0, 1]-clipped pre-training initialization masks W_i^p.

Claims & Grounded Evidence

4 Claims

Mechanism #r455-layer-split-elementwise-clipped-aggregation-mechanism

Combines layer-wise overwriting of the lower |Theta_i| - p generic layers (using fixed ones mask 1^{|Theta_i|-p}) with fine-grained element-wise interpolation of the top p higher layers via learnable weights W_i^p constrained to [0, 1] by element-wise clipping sigma(w) = max(0, min(1, w)) (Equation 4), and updates W_i^p via the chain-rule Hadamard gradient dL_i^t/dW_i^p = (dL_i^t/dTheta^hat_i^{t,p}) odot (Theta^{t-1} - Theta_i^{t-1})^p (Equation 5, Appendix A Equation 7) while freezing both Theta^{t-1} and Theta_i^{t-1}.

supports Source: FedALA: Adaptive Local Aggregation for Personalized Federated Learning · Locator: Section 3.2 Fine-Grained Model Aggregation & Weight Learning (Equations 2-5, Algorithm 1) and Appendix A Equation 7, pp. 11239-11240

Scope #r455-plug-and-play-personalized-initialization-scope

Operates purely as a client-side pre-training initialization step before local empirical risk minimization, requiring zero extra communication payload beyond standard global model broadcast/upload and functioning as a plug-and-play personalization module for existing federated optimizers (such as FedAvg, FedProx, andDitto).

supports Source: FedALA: Adaptive Local Aggregation for Personalized Federated Learning · Locator: Section 3.2 Discussion & Section 4.3 Applicability of ALA (Table 3), pp. 11240-11243

Limitation #r455-stateful-client-storage-and-warmup-cost-limitation

Requires clients to remain stateful across communication rounds to store their previous local model parameters Theta_i^{t-1} and persistent top-p aggregation weight tensor W_i^p, and incurs extra local gradient steps in the second round until W_i^p converges (typically ~10 epochs in round 2 followed by 1-2 epochs in subsequent rounds).

supports Source: FedALA: Adaptive Local Aggregation for Personalized Federated Learning · Locator: Section 3.2 Weight Learning & Appendix B Computation and Communication Overhead (Table 5), pp. 11240-11244

Application #r455-personalized-heterogeneous-vision-and-text-application

Applied to personalized federated image and text classification under pathological class-subset and Dirichlet label heterogeneity on MNIST, CIFAR-10, CIFAR-100, Tiny-ImageNet, and AG News.

supports Source: FedALA: Adaptive Local Aggregation for Personalized Federated Learning · Locator: Section 4 Experiments (Tables 1-4, Figures 3-5), pp. 11240-11243

⚠️ Curated Limitations & Negative Results

Requires stateful clients capable of persisting their previous local model Theta_i^{t-1} and top-p weight tensor W_i^p across communication rounds.
Incurs additional local weight-learning epochs during the initial ALA convergence phase in round 2.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	narrow_match	FedALA is a personalized federated learning algorithm that learns element-wise local-global model initialization weights for the top layers of each client model.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r455-primary-pdf`	FedALA: Adaptive Local Aggregation for Personalized Federated Learning Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Haibing Guan	primary paper full_text	ID: https://doi.org/10.1609/aaai.v37i9.26330 sha256:a94f10b7e1a0…
`r455-arxiv-abs`	FedALA: Adaptive Local Aggregation for Personalized Federated Learning (arXiv abstract) Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Haibing Guan	primary paper abstract_only	ID: https://arxiv.org/abs/2212.01197 sha256:cdc3282809a3…
`r455-metadata-cr`	FedALA: Adaptive Local Aggregation for Personalized Federated Learning (Crossref metadata) Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Haibing Guan	primary paper abstract_only	ID: https://doi.org/10.1609/aaai.v37i9.26330 sha256:0c417a7c2414…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T08:58:00Z

Updated: 2026-10-19T08:58:00Z

Source File: data/curated/FedAlaAdaptiveLocalAggregationForPersonalizedFederatedLearning.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