PaperlayerView exact revision on arXiv

Source-linked AI summary

Online Continual Learning in Edge-Cloud Applications

Charalampos Davalas, Dimitrios Michail, Christos Diou, Iraklis Varlamis, Konstantinos Tserpes

arXiv:2610.32472v1cs.LGcs.CV

TL;DR

Edge-deployed models need to adapt to non-stationary streams without the original training data while preserving learned knowledge. The paper integrates continual learning with ODIN- and ECDD-based sample selection in an edge-cloud pipeline; experiments across automotive, healthcare, and agricultural datasets support adaptation while mitigating catastrophic forgetting.

  • Problem

    Edge-deployed models must adapt to novel, non-stationary inputs without the original training dataset while retaining previously acquired knowledge.

    Source: §1 · p. 1

  • Method

    The paper proposes an edge-cloud continual-learning pipeline that uses ODIN confidence or ECDD drift detection to select informative samples and trigger rehearsal-based updates.

    Source: §1 · p. 2, §3.1.1 · p. 5, §3.1.2 · p. 6

  • Results

    Experiments across automotive, healthcare, and agricultural datasets show that selective sampling and rehearsal-based updates can support adaptation under distribution shifts while mitigating catastrophic forgetting.

    Source: §1 · p. 2, §7 · p. 16

  • Takeaways & Limitations

    The proposed edge-cloud pipeline provides an architectural setting for incorporating continual-learning and shift-detection mechanisms into IoT applications.

    Source: §7 · p. 16

    ODIN can be less scalable in larger architectures because of input-perturbation and temperature-scaling overhead, though the paper assumes lightweight IoT models make it more manageable.

    Source: §3.1.2 · p. 6

Abstract

from arXiv · show

Deployed machine learning models in real-world IoT and edge/cloud systems face a critical reliability bottleneck: the requirement for online adaptation to non-stationary data streams while operating without access to the full original training dataset. Performance is severely degraded when models encounter Out-of-Distribution (OOD) inputs or suffer from catastrophic forgetting. Addressing these challenges requires adaptation mechanisms capable of learning from evolving data streams while preserving previously acquired knowledge. This paper proposes an edge-cloud methodology for integrating Continual Learning (CL) with Out-of-Distribution (OOD) Detection and Concept Drift Detection in IoT applications. The methodology incorporates ODIN and ECDD as alternative mechanisms for identifying informative samples for annotation and triggering rehearsal-based model updates. Experimental results across automotive, healthcare, and agricultural datasets demonstrate the applicability of the proposed methodology and its potential to support model adaptation while mitigating catastrophic forgetting.

1 Introduction

Edge-deployed models must adapt to novel scenarios and evolving data without the original training set, while retaining previously acquired skills. This paper proposes an edge-cloud continual-learning pipeline that uses drift- or confidence-based sample selection to trigger rehearsal updates, evaluated across three IoT application domains.

1 Introduction

  • Deployed edge models face novel, out-of-distribution inputs without access to the original training data, risking performance degradation in mission-critical applications.The pipeline must identify novel scenarios, select samples for labeling, and integrate knowledge without losing prior skills.Source: §1 · p. 1
  • The proposed pipeline uses asynchronous, rehearsal-based fine-tuning triggered by conflict detection to prioritize informative samples for annotation and model updates.It is intended to support continual adaptation as data distributions evolve.Source: §1 · p. 2
  • The framework integrates out-of-distribution detection, concept-drift detection, and rehearsal-based continual learning in an edge-cloud deployment pipeline.The introduction presents this as a unified architecture for adaptive edge ML.Source: §1 · p. 2
  • Model-versus-user disagreement and model-confidence scoring are evaluated as alternative mechanisms for selecting informative samples in IoT environments.The selected samples support continual model adaptation.Source: §1 · p. 2
  • The paper evaluates adaptation to new stream data and retention of previously acquired knowledge across automotive, healthcare, and agricultural applications.The experiments examine the methodology's ability to adapt while mitigating catastrophic forgetting.Source: §1 · p. 2

2 Related Work

The related work spans IoT deployment, continual learning, online rehearsal, drift detection, and OOD detection. The paper positions its contribution as integrating streaming adaptation and distribution-shift detection into an edge-cloud architecture without predefined task boundaries.

2 Related Work

  • IoT and edge-cloud research addresses streaming data, limited access to original datasets, and operational constraints.The paper frames integration of continual-learning techniques in these systems as a developing research direction.Source: §2 · p. 2

2.1 Edge Computing in IoT

  • Edge-computing work addresses efficient deployment across diverse hardware, while RILOD incrementally builds datasets on-the-fly and EAR [2] uses shallow adaptors to mitigate catastrophic forgetting, demonstrating continual learning in edge–cloud systems.The work schedules pretrained models based on directed acyclic graphs and edge availability; its models are not necessarily trained continually. AdaInf schedules model training and inference based on model drift.Source: §2.1 · p. 2

2.2 Continual Learning and Catastrophic Forgetting

  • Traditional continual-learning methods often assume predefined task boundaries, whereas IoT environmental changes can be unpredictable.The paper extends these methods toward streaming adaptation without predefined boundaries.Source: §2.2 · p. 3

2.3 Online Continual Learning

  • Prior online continual-learning work uses rehearsal, including latent activations to avoid storing raw data and drift-triggered updates with rehearsal steps adjusted to drift magnitude.This paper builds on that foundation by integrating online learning into an edge-cloud architecture for IoT updates.Source: §2.3 · p. 3

2.4 Drift Detection Mechanisms

  • Drift detectors such as ADWIN and ECDD identify streaming distribution changes, while label-free model diagnostics may help when ground-truth labels are difficult to obtain.The paper cites its earlier work when motivating this label-free enhancement.Source: §2.4 · p. 3

2.5 Out-of-Distribution Detection

  • ODIN uses gradient-based input noise and softmax confidence to detect distribution shifts without modifying a pretrained model or requiring ground-truth labels.The paper adapts OOD detection to train on shifted samples rather than relying on arbitrary hyperparameters or randomized sampling.Source: §2.5 · pp. 3–4

3 Our Approach: An edge-cloud online continual learning pipeline

The paper presents an edge-cloud continual-learning pipeline that buffers selected stream samples for annotation and rehearsal-based model updates. ECDD and ODIN provide alternative detection mechanisms for selecting samples and guiding update timing.

3 Our Approach: An edge-cloud online continual learning pipeline

  • Novel or ambiguous edge samples are buffered for cloud annotation, used for continual-learning updates, and followed by redeployment to the edge.The framework uses incremental rehearsal-based updates rather than conventional full retraining on large training datasets.Source: Figure 1 · p. 4, §3 · pp. 4–5

3.1 User and Model Conflicts detection

  • Conflict-based storage uses driver–model disagreement or OoD detection, with detector signals and buffer conditions triggering updates instead of continuous training.The pipeline uses model classification statistics alongside detected drift or OoD signals.Source: §3.1 · p. 5

3.1.1 Drift detection

  • ECDD combines EWMA charts with a Bernoulli-based change detector over two streams and has stated O(1) complexity.Triggers include detected drift with a non-empty buffer, a full dynamic buffer, or a non-empty buffer after an idle period.Source: §3.1.1 · p. 5

3.1.2 Out of Distribution detection

  • ODIN scores confidence using temperature scaling and targeted input perturbations; its overhead may hinder scalability in larger architectures, though it is more manageable for lightweight IoT models.ODIN provides an OOD score without retraining the underlying model.Source: §3.1.2 · p. 6

3.2 Online continual learning on the cloud

  • Conflict-inducing samples are selected instead of reservoir-sampled, and the detection mechanism guides when buffered samples initiate rehearsal updates.If hardware permits, the setup can run on edge or cloud infrastructure, using a curated subset of stream data for adaptation.Source: §3.2 · p. 7

4 Experimental setup

The experiments evaluate classification accuracy in an edge-cloud continual-learning setup, comparing conflict-resolution and rehearsal combinations, data-specific accuracy measures, and model configurations.

4 Experimental setup

  • The framework tests classification accuracy across cloud-edge continual-learning interactions using sensor input.The experimental setup and implementations use TensorFlow 2.Source: §4 · p. 7

4.1 Method combinations

  • Method combinations pair conflict-resolution and rehearsal approaches; edge training has no sample selection, and asterisked methods require user input.Source: §4.1 · pp. 7–8, Table 2 · p. 8

4.2 Factory and Stream accuracy

  • Factory accuracy measures retained knowledge on held-out pretraining samples, while Stream accuracy measures accuracy on held-out samples originating from the stream.Both use the original classification-accuracy metric; they differ in the data used.Source: §4.2 · p. 8

4.3 Models in use

  • Except in the toy scenario, experiments use MobileNet and begin online training with a factory pre-trained model trained on a preset sample amount that cannot cover every real-world variation.Source: §4.3 · pp. 8–9

5 Assisted driving

The assisted-driving experiments examine continual adaptation to changing traffic-sign conditions and road-object streams. Across these settings, rehearsal with selective sampling supports accuracy gains while preserving factory performance.

5 Assisted driving

  • Traffic-sign recognition remains challenging across countries and local conditions, while dataset augmentation or expansion can improve coverage but become complex and costly over time.The experiment uses a customized Belgium Traffic Signs dataset with environmental and equipment distortions.Source: §5 · p. 9, §5.1 · p. 9

5.1 Training with driver reactions to traffic signs

  • The traffic-sign proof of concept maps signs to stopping, decelerating, or maintaining speed, with driver reactions recorded alongside model predictions.Drivers see variations of a 22,400-sample stream, and the assumed driver accuracy is 95%.Source: §5.1 · p. 9
  • ODIN increases stream accuracy faster while maintaining both stream and factory accuracy throughout the traffic-sign experiment.The initial factory and stream accuracies are 0.9533 and 0.7125, respectively.Source: §5.1 · p. 9, Figure 2 · p. 10

5.2 Training for general classification on the road

  • The nuScenes stream uses detected objects from the dataset, restricted to its ten most-sampled classes because class imbalance prevents controlled class-ratio sampling.Training, test, factory-accuracy, and stream-accuracy samples are disjoint.Source: §5.2 · p. 11, Table 4 · p. 12
  • ANDE+ODIN is the most capable nuScenes combination for both factory and stream accuracy, and ODIN maintains accuracy consistently while random sampling causes high deviations.The comparison is reported in Table 5 and Figure 3.Source: Figure 3 · p. 11, §5.2 · p. 12, Table 5 · p. 12

6 Infection detection

The infection-detection scenarios consider continual learning as support for medical and agricultural professionals, with human input retained. Results differ by task: rehearsal methods perform well on malaria classification, while selective methods mitigate forgetting in crop classification.

6 Infection detection

  • The medical scenario uses an edge device such as a smart microscope for on-site cell-image classification, as a supplementary tool rather than a replacement for professional judgment.Professionals may correct conflicts; uncertain predictions can also be referred for qualified annotation or user override.Source: §6 · p. 12, §6.1 · pp. 13–14

6.1 Infection detection with cell image classification

  • On the relatively simple binary malaria task, catastrophic forgetting is not observed, and rehearsal approaches maintain high accuracy with minimal human intervention.ECDD and ODIN combinations show particularly strong overall performance, though per-timestep deviation remains a concern.Source: Figure 4 · p. 13, §6.1 · p. 14, Table 6 · p. 14

6.2 Crop inspection with leaf samples classification

  • The crop scenario targets early disease detection from drone images, with expert feedback when the model’s decision is uncertain.The experiment uses the TensorFlow Plant Village dataset and similar augmentations to the malaria example.Source: §6.2 · p. 14
  • RH+ODIN and RH+ECDD adequately mitigate the catastrophic forgetting induced by the TRN baseline on Plant Village, increasing average factory and stream accuracy.The reported comparison is shown in Figure 5 and Table 7.Source: §6.2 · pp. 14, 16, Table 7 · p. 14, Figure 5 · p. 15

7 Discussion

The discussion presents rehearsal-based continual learning as a way to support adaptation under distribution shifts while mitigating catastrophic forgetting in the evaluated datasets. It also emphasizes that method suitability varies by application.

7 Discussion

  • Experiments indicate that selective sampling and rehearsal-based updates support adaptation under distribution shifts while mitigating catastrophic forgetting in the evaluated datasets.The framework is presented for cloud-edge scenarios including assisted driving and supplementary infection detection.Source: §7 · p. 16
  • Combining rehearsal with drift detection or OoD detection can support both forgetting mitigation and evolution of a pre-trained model’s classification ability.The discussion states that no single approach adapts to every scenario.Source: §7 · p. 16
  • Sample storage may require user permission, and the authors note that the methods could be further improved.Source: §7 · p. 16

Related papers

Loading 2610.32472v1…