PaperlayerView exact revision on arXiv

Source-linked AI summary

Sylvas: Synergistic Learning Value based Device Scheduling in Federated Continual Learning

Yuxuan Sun, Yuxuan Bai, Tan Chen, Sheng Zhou, Zhisheng Niu

arXiv:2609.15763v1cs.LGcs.AI

TL;DR

FCL must adapt shared models to evolving distributed data while retaining historical knowledge, but device scheduling is difficult under distribution shifts, label scarcity, and edge-resource constraints. Sylvas combines distributional value with pseudo-label quantity and reliability to schedule high-value devices. Case studies report timely adaptation and effective use of unlabeled data.

  • Problem

    FCL scheduling must quantify device contributions under spatio-temporal distribution dynamics and label scarcity while meeting resource constraints for timely model updates.

    Source: §I · p. 1

  • Method

    Sylvas integrates temporal drift and collective divergence with pseudo-label quantity and reliability into a synergistic value used to schedule devices under resource constraints.

    Source: §I · p. 1, §III · p. 3

  • Results

    Case studies show Sylvas recovers more rapidly after distribution changes, maintains more stable performance across continual-learning stages, and achieves higher test accuracy while better exploiting unlabeled data.

    Source: A. Sylvas with Fully Labeled Data · p. 6, B. Sylvas in Semi-Supervised FCL · p. 7

  • Takeaways & Limitations

    Balancing pseudo-label quantity and reliability improves model performance with limited labeled data and the timeliness of FCL.

    Source: B. Sylvas in Semi-Supervised FCL · p. 7

Abstract

from arXiv · show

Federated continual learning (FCL) enables shared global models to continuously adapt to distributed and non-stationary data streams, making it important for Internet of Things applications such as intelligent transportation, industrial monitoring, and unmanned systems. Under spatio-temporal data distribution dynamics and label scarcity, a key challenge is how to quantify the contribution of each edge device to global learning performance and schedule the most valuable devices under resource constraints for timely model updating. This article presents Sylvas, a synergistic learning value based device scheduling framework for FCL at the wireless edge. Sylvas evaluates the learning value of distributed data from two perspectives: distributional value, which characterizes the contribution of device data to global model learning from a spatio-temporal distribution perspective, and label value, which captures the quantity and reliability tradeoff of pseudo-labeled data. By integrating these factors into a synergistic learning value metric, Sylvas schedules devices with high learning value while satisfying communication and computation resource constraints. Case studies demonstrate that Sylvas supports timely model adaptation under spatio-temporal distribution dynamics and effectively exploits unlabeled data.

I. INTRODUCTION

FCL addresses continuously evolving data streams by adapting a shared model while retaining historical knowledge. Sylvas schedules devices by combining spatio-temporal distributional value and pseudo-label quantity-reliability value under resource constraints.

I. INTRODUCTION

  • FCL must adapt to evolving data streams while retaining historical knowledge, and timely updates aim to maximize performance within latency budgets.Overemphasizing new data may degrade performance on historical data and lead to catastrophic forgetting.Source: §I · p. 1
  • Limited communication bandwidth and computation make it impractical to involve every edge device in each training round, motivating value-aware device scheduling.Device value depends on spatio-temporal representation and, for unlabeled data, pseudo-label quantity and reliability.Source: §I · p. 1
  • Sylvas combines temporal drift and collective divergence as distributional value, and balances pseudo-label quantity against reliability as label value.It integrates these perspectives into a synergistic learning value for scheduling high-value devices under communication and computation constraints.Source: §I · p. 1
  • Case studies demonstrate timely model adaptation under spatio-temporal distribution dynamics and effective use of unlabeled data in semi-supervised FCL.When all device data are labeled, Sylvas reduces to a case using distributional value alone.Source: §I · p. 2

II. FCL AT THE WIRELESS EDGE: SYSTEM MODEL AND DEVICE SCHEDULING CHALLENGES

At the wireless edge, devices stream data and update local models while an edge server coordinates aggregation into a shared global model. FCL seeks timely adaptation within a latency budget while retaining historical knowledge, so scheduling must choose a resource-feasible subset of devices.

II. FCL AT THE WIRELESS EDGE: SYSTEM MODEL AND DEVICE SCHEDULING CHALLENGES

  • Devices stream data and update local models, whose updates are aggregated under an edge server to train a shared global model.New samples may be labeled or unlabeled; in the semi-supervised setting, the server retains a small labeled reference dataset for initialization, calibration, and evaluation.Source: §II · p. 2
  • FCL seeks to maximize learning performance within a latency budget or minimize the latency needed to reach a target performance.The target must reflect both adaptation to new data and retention of historical knowledge.Source: §II · p. 2
  • Because spatio-temporal dynamics, label scarcity, and edge-resource limits make it impractical to schedule all devices each round, FCL must select a high-contribution subset under resource constraints.The scheduling objective is timely model adaptation.Source: §II · p. 2

A. Spatio-Temporal Distribution Dynamics

Device data distributions differ and evolve over time, so a device’s learning contribution depends both on new data’s deviation from historical data and on how scheduled devices collectively represent the global distribution.

A. Spatio-Temporal Distribution Dynamics

  • Temporal drift captures how newly arriving data increasingly deviate from historical data, reflecting new knowledge for the global model.Spatial contribution depends on how a device’s data complement other scheduled devices and how well the group represents the global distribution.Source: A. Spatio-Temporal Distribution Dynamics · p. 2
  • The distributional value metric is motivated by the joint temporal and spatial dimensions of device-data contribution.Temporal drift and collective representation together characterize the spatio-temporal distribution dynamics relevant to scheduling.Source: A. Spatio-Temporal Distribution Dynamics · pp. 2–3

B. Label Scarcity

Label scarcity makes pseudo-label quantity and reliability jointly relevant to device learning value, while resource constraints require balancing learning contribution against communication and computation costs. Sylvas integrates distributional and label value with device capabilities and wireless conditions for scheduling.

C. Resource Constrained Device Scheduling

  • Scheduling more devices may accelerate convergence but increase delay, whereas selecting only fast devices may reduce delay but limit learning performance.Under synchronous aggregation, a round waits for all scheduled devices to finish local computation and transmission.Source: C. Resource Constrained Device… · p. 3

III. SYLVAS: SYNERGISTIC LEARNING VALUE BASED DEVICE SCHEDULING

  • Sylvas integrates distributional value and label value with device computing capabilities and wireless conditions to schedule high-value devices under resource and latency constraints.Distributional value captures spatio-temporal data contribution; label value balances pseudo-label quantity and reliability.Source: §III · p. 3

A. Distributional Value for Spatio-Temporal Dynamics

Sylvas defines distributional value by combining device-level temporal drift with group-level collective divergence. Its scheduling objective shifts emphasis from adapting to new data toward representative, stable device groups as training proceeds.

A. Distributional Value for Spatio-Temporal Dynamics

  • Collective divergence measures the discrepancy between scheduled devices’ aggregate data distribution and the global distribution.Complementary device data can make the group aggregate closer to the global distribution even when individual devices are biased.Source: A. Distributional Value for… · p. 3
  • Temporal drift measures the discrepancy between each device’s current and historical data distributions.It is characterized through changes in class proportions for class-incremental settings and changes in average model updates for domain-incremental settings.Source: A. Distributional Value for… · p. 3
  • Sylvas rewards high temporal drift and penalizes high collective divergence, shifting from adaptation early to representativeness and stability later.Devices train on newly arrived data alongside a limited set of representative historical samples.Source: A. Distributional Value for… · p. 3

B. Label Value for Unlabeled Data

Sylvas uses label value to balance the quantity and reliability of pseudo-labeled data. Its semi-supervised workflow filters device-generated pseudo-labels and adapts confidence thresholds for newly introduced classes.

B. Label Value for Unlabeled Data

  • Devices generate pseudo-labels from the global model and retain samples above a confidence threshold for local training.The server initializes the model with a small labeled reference dataset, aggregates device updates, and then fine-tunes or calibrates the model with that reference data.Source: B. Label Value for Unlabeled Data · p. 4
  • Label value balances the number of unlabeled samples passing a confidence threshold against the average confidence of retained pseudo-labels for each class.More retained samples provide richer information for updating, while higher reliability reduces adverse effects from incorrect pseudo-labels.Source: B. Label Value for Unlabeled Data · p. 4
  • Sylvas begins with a relaxed threshold for newly introduced classes and raises it as predictions become more reliable, while keeping stricter filtering for previously learned classes.This threshold schedule is intended to retain useful new-class samples early and suppress incorrect pseudo-labels as training proceeds.Source: B. Label Value for Unlabeled Data · p. 4

C. Sylvas under Resource Constraints

Sylvas schedules feasible device groups by balancing synergistic learning value against computation, communication, and latency constraints. Its group-level objective makes selection a combinatorial problem rather than an independent ranking of devices.

C. Sylvas under Resource Constraints

  • High synergistic learning value can come with higher training latency, because the slowest scheduled device may determine synchronous round completion time.Selecting more high-value devices may improve global-model accuracy and reduce rounds to convergence, but may also increase per-round latency.Source: C. Sylvas under Resource Constraints · p. 4
  • Each round, Sylvas selects a feasible device group with high synergistic learning value while meeting prescribed latency and resource constraints.Because collective divergence depends on the selected group’s composition, devices cannot be chosen by independently ranking individual values.Source: C. Sylvas under Resource Constraints · p. 4
  • Sylvas uses fixed-sum coordinate descent to exchange scheduled and unscheduled devices, and a greedy strategy to add devices with the largest feasible marginal gain.The coordinate-descent strategy evaluates candidate group sizes; the greedy strategy respects latency and resource constraints.Source: C. Sylvas under Resource Constraints · p. 5

IV. CASE STUDIES

The case studies consider fully labeled FCL using distributional value alone and semi-supervised FCL incorporating label value to exploit device-side unlabeled data.

IV. CASE STUDIES

  • The evaluation covers fully labeled FCL with distributional value alone and semi-supervised FCL with label value added for device-side unlabeled data.The semi-supervised setting uses pseudo-label quantity and reliability.Source: §IV · p. 5

A. Sylvas with Fully Labeled Data

In fully labeled class-incremental FCL, Sylvas combines temporal drift and collective divergence to schedule devices as data distributions change. Its case study reports faster recovery after distribution changes and more stable performance across continual-learning stages.

A. Sylvas with Fully Labeled Data

  • The CIFAR-100 evaluation uses 30 devices, ResNet-18, successive class additions, and a 30-second per-round latency budget.Twelve devices receive samples from newly introduced classes at each incremental stage.Source: A. Sylvas with Fully Labeled Data · p. 5
  • The temporal-drift weight starts at 2.0 and decreases linearly to zero over each incremental stage's communication rounds.This shifts emphasis from newly arrived classes early in a stage toward the selected group’s global representativeness later.Source: A. Sylvas with Fully Labeled Data · p. 5
  • The comparison includes seven baselines, including collective-divergence-only FedCGD and temporal-drift-only Pure Drift.Other baselines include random, channel-based, loss-based, update-norm, and class-imbalance scheduling.Source: A. Sylvas with Fully Labeled Data · p. 5
  • Sylvas recovers more rapidly after each distribution change and maintains more stable performance by jointly considering temporal drift and collective divergence.The evaluation plots round-by-round accuracy and the best accuracy achieved up to each communication round.Source: A. Sylvas with Fully Labeled Data · pp. 5–6, Figure 4 · p. 6

B. Sylvas in Semi-Supervised FCL

In semi-supervised FCL, Sylvas schedules devices using the quantity and reliability of their pseudo-labeled data. The case study reports fewer rounds to target accuracies and higher test accuracy, with fewer rounds reducing adaptation latency under the shared per-round budget.

B. Sylvas in Semi-Supervised FCL

  • Candidate devices report pseudo-label quantity and reliability statistics; the server schedules a subset, whose retained pseudo-labeled samples are used for local training.After aggregation, the server fine-tunes the global model using labeled reference data.Source: B. Sylvas in Semi-Supervised FCL · p. 6
  • Sylvas raises the confidence threshold for newly introduced classes from 0.5 to 0.8 while keeping the threshold for previously learned classes at 0.95.Samples meeting neither threshold are excluded from the unsupervised loss.Source: B. Sylvas in Semi-Supervised FCL · p. 6
  • Comparisons include new-class-only, all-device, and no-device training, plus FL2, FedLGMatch, and FSSL-UC.The setting uses SVHN, 30 devices, limited labeled server data, and a one-second per-round latency budget.Source: B. Sylvas in Semi-Supervised FCL · p. 6
  • Sylvas requires fewer communication rounds to reach target accuracies and achieves higher test accuracy; under the shared per-round latency budget, fewer rounds mean lower adaptation latency.The advantage in rounds is more pronounced at higher accuracy targets.Source: B. Sylvas in Semi-Supervised FCL · pp. 6–7, Figure 5 · p. 6

V. CONCLUSION AND OUTLOOK

The conclusion presents Sylvas as a resource-constrained device-scheduling framework that integrates distributional and label value. Case studies report improved learning performance, fewer adaptation rounds, and effective use of unlabeled data; the outlook identifies mobility, annotation, and training-triggering questions.

V. CONCLUSION AND OUTLOOK

  • Case studies in fully labeled and semi-supervised FCL show improved learning performance under evolving distributions, fewer adaptation rounds, and effective use of unlabeled data.Scheduling integrates synergistic learning value with communication and computation costs.Source: §V · p. 7
  • Future mobility-aware FCL work should jointly model mobility, data drift, and communication opportunities.The outlook concerns dynamic edge systems with vehicles, drones, and mobile robots.Source: §V · p. 7
  • The paper proposes integrating active learning to select valuable or uncertain samples for annotation and reduce labeling costs.The stated aim also includes improving new-class learning.Source: §V · p. 7
  • Future resource-efficient continual learning could trigger training when significant data drift or performance degradation is detected rather than at fixed intervals.Reliable triggering policies under privacy and resource constraints remain an open problem.Source: §V · p. 7
Loading 2609.15763v1…