Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Neuro-Symbolic Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedNSL, a federated neuro-symbolic learning framework that couples server-side prior rule distribution learning with client-side posterior rule learning via distribution-coupled bilevel optimization and a factorized federated variational EM with a KL constraint.
Problem
Neuro-symbolic learning requires access to downstream data, which is infeasible under federated learning; rule distributions are heterogeneous across clients and the rule search space is huge.
Approach
Replace weight-coupled personalization with a distribution-coupled bilevel formulation in which latent rule distributions (not weights) are the FL communication medium; a KL constraint aligns client posteriors to the server prior, and variational EM reduces the rule search space.
Key results
- Outperforms the best baseline by 17% in unbalanced average training accuracy and 29% in unseen average testing accuracy (abstract claim).
- On DWIE, groups with KL-divergence constraints reach higher F1 and logic accuracy than groups without the constraint (qualitative figure comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a KL-divergence rule-distribution constraint (ablation) | FedNSL · DWIE seen and unseen testing data | F1-score | Helped |
| Add a KL-divergence rule-distribution constraint (ablation) | FedNSL · DWIE | logic accuracy | Helped |
| Remove the KL-divergence rule-distribution constraint (ablation) | FedNSL · DWIE seen testing data | F1-score convergence | Hurt |
| Remove the KL-divergence rule-distribution constraint (ablation) | FedNSL · DWIE unseen testing data | F1-score | Hurt |
| Increase the degree of training-testing data heterogeneity | FedNSL versus baselines · Synthetic numerical experiment | average training and unseen testing accuracy | Helped |
Limitations
- Real-data evaluation is limited to one dataset (DWIE) and four clients.
- Results are largely shown in figures; numeric values for the headline gains are only given in the abstract.
Experimental setup
Datasets: DWIE synthetic 2D Gaussian data
Benchmarks: synthetic 3-class federated classification DWIE relation prediction
Metrics: training accuracy unseen testing accuracy F1-score logic accuracy
Baselines: pFedMe Per-FedAvg pFedBayes LR-XFL FedU
Builds on: variational expectation maximization pFedBayes FedEM LR-XFL
Closest papers by meaning
- Convex and Bilevel Optimization for Neural-Symbolic Inference and Learning (ICML 2024 · similarity 0.85)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.85)
- FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data (ICLR 2024 · similarity 0.85)
- Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models (NeurIPS 2026 · similarity 0.84)
- Position: The Road to Generalizable Neuro-Symbolic Learning Should be Paved with Foundation Models (NeurIPS 2026 · similarity 0.84)
- DOLPHIN: A Programmable Framework for Scalable Neurosymbolic Learning (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.