Home / Curated / Algorithms / FedBABU (Federated Averaging with Body Aggregation and Body Update)

FedBABU (Federated Averaging with Body Aggregation and Body Update)

Algorithms AIMech:FedBabuFederatedAveragingWithBodyAggregationAndBodyUpdate DRAFT

Aliases: FedBABU, Federated Averaging with Body Aggregation and Body Update, Frozen Orthogonal-Head Body-Only Federated Representation Learning

A decoupled representation-learning federated optimization algorithm (Oh, Kim, & Yun 2022, Section 5, Algorithms 1-2) that decomposes the deep network parameters theta = {theta_ext, theta_cls} into a feature extractor body theta_ext and a linear classification head theta_cls in R^{C x d}, freezes theta_cls = theta_{G,cls}^0 at its shared random orthogonal initialization (learning rate alpha_cls = 0) throughout all federated training rounds so clients update and server-aggregate only the body theta_ext against fixed shared class templates, and subsequently adapts the model to each client via local fine-tuning.

Identity Notes & Concept Scope Distinct from FedRep (AIMech:FedRepExploitingSharedRepresentationsForPersonalizedFederatedLearning, Collins et al. 2021), which actively updates each client's local linear head theta_{i,cls} via local gradient steps before updating the shared body in every round, whereas FedBABU keeps a single shared random orthogonal head completely frozen (alpha_cls = 0) during federated representation learning so all clients share identical fixed class decision boundaries.

Claims & Grounded Evidence

4 Claims

Mechanism #r424-head-update-degradation-of-personalization-and-shared-orthogonal-template-mechanism

Jointly updating and aggregating the classifier head theta_cls under non-IID client label distributions (as in FedAvg) causes divergent local head updates for missing classes that degrade the feature extractor's personalization capability (Section 3, Tables 1-3); freezing theta_cls at a shared orthogonal initialization forces all clients to map each class to the same fixed row template of theta_{G,cls}^0 without inter-client head interference (Section 4-5).

supports Source: FedBABU: Toward Enhanced Representation for Federated Image Classification · Locator: Section 3-4 Empirical Observations on Decoupled Parameter Updates (Tables 1-5, Figures 1-2) & Section 5 FedBABU

Mechanism #r424-single-epoch-head-or-full-fine-tuning-rapid-personalization

Because the body theta_{G,ext}^T learned with a frozen orthogonal head already produces well-clustered class representations, fine-tuning each client's local model after federated training converges within 1 to 5 local epochs and outperforms personalized baselines including FedPer, LG-FedAvg, FedRep, Per-FedAvg, andDitto (Section 5.2, Tables 6-8).

supports Source: FedBABU: Toward Enhanced Representation for Federated Image Classification · Locator: Section 5.2 Experimental Results (Tables 6-9, Figure 3), pp. 6-9

Limitation #r424-classification-head-architecture-and-local-finetuning-data-limitation

Restricted to supervised classification architectures with an explicit final linear classifier layer theta_cls, and personalized deployment requires labeled local data on each target client to perform the post-training fine-tuning step (Algorithm 2).

supports Source: FedBABU: Toward Enhanced Representation for Federated Image Classification · Locator: Section 5 Algorithm 1-2 & Appendix B-C

Application #r424-cifar10-cifar100-and-shard-partitioned-mobile-net-resnet-application

Applied to federated image classification on shard-partitioned CIFAR-10 and CIFAR-100 with ConvNet, MobileNet, ResNet, and VGG backbones under varying client participation rates and local epoch budgets.

supports Source: FedBABU: Toward Enhanced Representation for Federated Image Classification · Locator: Section 5.2 & Appendix D-I (Tables 6-9, Tables 13-22)

⚠️ Curated Limitations & Negative Results

Requires feature dimension d >= number of classes C to initialize exact pairwise orthogonal class rows in theta_{G,cls}^0 (otherwise relies on random near-orthogonal rows).
Does not train a personalized head during the federated rounds themselves, requiring a post-training local fine-tuning pass on each client.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	FedBABU is a body-only federated representation learning and personalization algorithm within Federated Learning.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r424-primary-pdf`	FedBABU: Toward Enhanced Representation for Federated Image Classification Jaehoon Oh, SangMook Kim, Se-Young Yun	primary paper full_text	ID: https://openreview.net/forum?id=HuaYQfggn5u sha256:02f54062a558…
`r424-arxiv-abs`	FedBABU: Toward Enhanced Representation for Federated Image Classification (arXiv abstract) Jaehoon Oh, SangMook Kim, Se-Young Yun	primary paper abstract_only	ID: https://arxiv.org/abs/2106.06042 sha256:dfea86449d92…
`r424-metadata-cr`	FedBABU: Toward Enhanced Representation for Federated Image Classification (Crossref metadata) Jaehoon Oh, SangMook Kim, Se-Young Yun	primary paper abstract_only	ID: https://openreview.net/forum?id=HuaYQfggn5u sha256:b0442817d33f…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T08:27:00Z

Updated: 2026-10-19T08:27:00Z

Source File: data/curated/FedBabuFederatedAveragingWithBodyAggregationAndBodyUpdate.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