Home / Curated / Algorithms / GLFC (Global-Local Forgetting Compensation)

GLFC (Global-Local Forgetting Compensation)

Algorithms AIMech:GlfcGlobalLocalForgettingCompensationForFederatedClassIncrementalLearning DRAFT

Aliases: GLFC, Global-Local Forgetting Compensation, Global-Local Forgetting Compensation for Federated Class-Incremental Learning

GLFC (Global-Local Forgetting Compensation) is a federated class-incremental learning (FCIL) algorithm that mitigates both local catastrophic forgetting caused by client class imbalance between old and new classes and global catastrophic forgetting caused by non-i.i.d. cross-client class distributions by combining (1) a local class-aware gradient compensation loss L_GC that normalizes per-sample binary cross-entropy by separate old-class and new-class average gradient magnitudes (G_o and G_n), (2) a class-semantic relation distillation loss L_RD that aligns current predictions P_l^t(X_{lb}^t, Theta_l^{r,t}) with hybrid targets Y_l^t(X_{lb}^t, Theta_l^{t-1}) incorporating soft old-class probabilities from the globally selected best old model Theta^{t-1}, and (3) a proxy server S_P that detects task transitions via local prediction entropy jumps and selects Theta^{t-1} using perturbed prototype gradients nabla Gamma_{lc}.

Identity Notes & Concept Scope Introduced by Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu in 'Federated Class-Incremental Learning' (CVPR 2022, pp. 10154-10163, DOI:10.1109/CVPR52688.2022.00992; arXiv:2203.11473). Distinct from FedWeIT (Yoon et al. ICML 2021), which addresses task-incremental federated continual learning with known task IDs at test time and parameter decomposition into masked base weights and sparse task-adaptive weights rather than single-head class-incremental learning; and distinct from centralized class-incremental learning methods (iCaRL, Rebuffi et al. CVPR 2017; PODNet, Douillard et al. ECCV 2020), which assume a single centralized exemplar store without heterogeneous client groups S_o, S_b, S_n or proxy-server prototype gradient reconstruction.

Claims & Grounded Evidence

4 Claims

Mechanism #glfc-class-aware-gradient-compensation-loss

For a mini-batch {(x_{li}^t, y_{li}^t)}_{i=1}^b on client S_l at incremental task t, GLFC measures the output-layer gradient scalar G_{li}^t = P_l^t(x_{li}^t, Theta_l^{r,t})_{y_{li}^t} - 1 (Equation 2), computes separate mean gradient magnitudes G_n for new classes y_{li}^t in Y_l^t and G_o for old classes y_{li}^t in U_{j=1}^{t-1} Y_l^j (Equation 3), and reweights binary cross-entropy via L_GC = (1/b) sum_{i=1}^b (|G_{li}^t| / bar_G_i) D_CE(P_l^t(x_{li}^t, Theta_l^{r,t}), y_{li}^t) (Equation 4), equalizing the aggregate gradient update magnitude between minority old exemplars and majority new-class samples.

supports Source: Federated Class-Incremental Learning · Locator: Section 4.1, Equations (2)-(4), page 4

Mechanism #glfc-class-semantic-relation-distillation-loss

To preserve inter-class semantic relations across old classes C^p = sum_{i=1}^{t-1} C^i and new classes C^t, GLFC constructs hybrid targets Y_l^t(X_{lb}^t, Theta_l^{t-1}) in R^{b x (C^p + C^t)} by replacing the first C^p dimensions of one-hot ground-truth labels Y_{lb}^t with soft probabilities P_l^{t-1}(X_{lb}^t, Theta_l^{t-1}) predicted by the previous task model Theta_l^{t-1}, and minimizes the Kullback-Leibler divergence L_RD = D_KL(P_l^t(X_{lb}^t, Theta_l^{r,t}) || Y_l^t(X_{lb}^t, Theta_l^{t-1})) (Equation 5) jointly with L_GC in L_l = lambda_1 L_GC + lambda_2 L_RD (Equation 6).

supports Source: Federated Class-Incremental Learning · Locator: Section 4.1, Equations (5)-(6), page 4

Mechanism #glfc-entropy-jump-detection-and-proxy-server-prototype-gradient-selection

Clients detect the arrival of new classes without prior task-transition knowledge when the increase in mean prediction entropy H_l^{r,t} = (1/|X_l^t|) sum_i H(P_l^t(x_{li}^t, Theta_l^{r,t})) satisfies H_l^{r,t} - H_l^{r-1,t} >= r_h = 1.2 (Equation 7); upon transition, each client perturbs class-mean prototypes x^p_{lc} via latent Gaussian noise backpropagation (Equations 10-11) and transmits shallow network gradients nabla Gamma_{lc} (Equation 8) to the proxy server S_P, which reconstructs prototype images via L-BFGS (Equation 9) and evaluates candidate client models across all reconstructed classes to broadcast the best old model Theta^{t-1} for relation distillation.

supports Source: Federated Class-Incremental Learning · Locator: Section 4.2, Equations (7)-(11), pages 5-6; Supplementary Algorithm 1, page 12

Application #glfc-fcil-benchmark-gains-on-cifar100-imagenet-subset-and-tiny-imagenet

On federated class-incremental benchmarks with 10 incremental tasks and non-i.i.d. client streams (including dynamically arriving clients in S_o, S_b, and S_n), GLFC improves average top-1 accuracy by 6.0%-14.1% on CIFAR-100, 7.6%-13.9% on ImageNet-Subset, and 5.3%-11.6% on Tiny-ImageNet over federated adaptations of iCaRL, BiC, PODNet, DDE+iCaRL, GeoDL+iCaRL, and SS-IL (Tables 1-3).

supports Source: Federated Class-Incremental Learning · Locator: Section 5.2, Tables 1-3, pages 6-8

⚠️ Curated Limitations & Negative Results

Requires each client S_l to maintain a bounded local exemplar memory M_l (e.g., 2,000 samples divided across seen classes via herding) to supply old-class samples for L_GC; when exemplar memory is completely disabled (GLFC-w/oMem), CIFAR-100 average accuracy drops by 10.6%-12.0%.
Proxy-server selection of the best historical model Theta^{t-1} requires L-BFGS gradient inversion over shallow encoder gradients nabla Gamma_{lc} whenever a task transition is detected.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	GLFC is a specific federated class-incremental learning algorithm that compensates for local and global catastrophic forgetting across distributed clients under AIO:FederatedLearning.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r515-primary-pdf`	Federated Class-Incremental Learning Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, Qi Zhu	primary paper full_text	ID: https://dblp.org/rec/conf/cvpr/DongWFSXW022 sha256:fde0980a2a5e…
`r515-arxiv-abs`	Federated Class-Incremental Learning (arXiv abstract) Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, Qi Zhu	primary paper abstract_only	ID: https://arxiv.org/abs/2203.11473 sha256:f47c3caff455…
`r515-metadata-cr`	Federated Class-Incremental Learning (Crossref metadata) Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, Qi Zhu	primary paper abstract_only	ID: https://dblp.org/rec/conf/cvpr/DongWFSXW022 sha256:8e713776ba72…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:58:00Z

Updated: 2026-10-19T09:58:00Z

Source File: data/curated/GlfcGlobalLocalForgettingCompensationForFederatedClassIncrementalLearning.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