A lightweight machine-learning method reported less forgetting and higher final accuracy than the FedAvg baseline in a simulated test of changing artistic styles, according to a preprint on cultural-heritage AI. FedCurv-DR reached 66.92% final stream accuracy, compared with 63.18% for FedAvg; its backward-transfer (BWT) score was -6.92 percentage points versus -11.30 points. BWT is the study’s measure of how well earlier learning is retained as new experiences arrive, so the less-negative score indicates better retention in this benchmark.
FedCurv-DR is a lightweight, regularization-based strategy—in effect, it adds a penalty around parameters judged important to earlier learning. It accumulates those importance estimates across clients and experiences, updates them only at fixed intervals and does not use a replay buffer; the paper calls the approach exemplar-free.
The study asked whether this design could balance final accuracy, retention, differences in accuracy between clients and energy use.
When the artwork keeps changing
The classification goal stayed fixed: identify genres. Artistic styles were introduced one after another as domain-incremental experiences, creating a simulated federated continual-learning setting in which the model learned from a changing stream.
Researchers used a selected subset of refined WikiArt containing approximately 20,000 images. The dataset has 11 genres and 27 styles; after two genres were dropped, the experiment retained nine genres and selected four style experiences.
Within each experience, the images were divided among three non-IID clients using a Dirichlet distribution, with 20% of each partition kept for local validation. The simulated clients therefore did not receive identical mixes of material.
Retention came with an energy trade-off
Researchers compared FedAvg with original FedCurv using EWC and EWC-DR importance estimates, plus two FedCurv-DR versions that refreshed estimates at different intervals. The reported FedCurv-DR settings used a decay factor of 0.95.
With an update interval of two experiences, FedCurv-DR had the best listed final-stream results: 66.92% accuracy and BWT of -6.92 percentage points. At interval five, it recorded 65.86% accuracy and -7.49 points. Both FedCurv-DR versions were higher on accuracy and less negative on BWT than FedAvg.
The energy picture was different. FedAvg used 0.0296 kWh, the lowest measured amount. FedCurv-DR at interval five used 0.0334 kWh, less than interval two’s 0.0385 kWh, while original FedCurv with EWC and EWC-DR used 0.0448 and 0.0458 kWh.
The figures show a trade-off in this setup: interval five used less energy than interval two, but its final accuracy was lower. The study tracked energy use and equivalent carbon emissions with CodeCarbon.
A fairness signal, not a verdict
The study treated the standard deviation of client accuracies as a proxy for client-level fairness. On that measure, FedAvg’s final disparity was 7.22 percentage points, compared with 2.89 points for FedCurv-DR at interval two and 3.09 for original FedCurv with EWC.
Those figures are a signal about the simulated partitions, not a finding that the system is fair in real-world use. The authors call the fairness results inconclusive, and the clients were synthetic rather than real institutions.
The paper describes current-experience accuracy as competitive, but supplies no numerical values for that measure in the reported text.
The test stops short of deployment
The runs used a pretrained EfficientNet-B0 backbone with approximately 5.3 million parameters and took place on one machine with an 8GB NVIDIA RTX 4070 GPU. All methods used the same training settings, including two local epochs per round, five rounds per experience and 20 rounds overall.
That leaves the results as evidence from a controlled, single-machine simulation using a selected WikiArt subset and synthetically partitioned clients. The experiment does not show how the method would perform in real cultural-heritage institutions or deployed federated systems.
The reported analysis includes no confidence intervals, p-values or repeated-run variability, making it difficult to judge how stable the differences would be across runs.
The paper identifies larger-scale tests with real distributed clients and ablation studies of the update interval, decay factor and regularisation strength as next steps. It also recommends secure aggregation and differential privacy for deployment, but those protections were not evaluated in the reported experiments.
The document is listed as an arXiv preprint dated 20 August 2026 and identified in its front matter as an accepted author manuscript for IEEE-CH 2026.
Paper data and sources
Original title: An Inclusive and Lightweight Approach to Federated Continual Learning for Cultural Heritage
Authors: Ioannis Theologitis, Debin Meng, Stylianos Eleftheriadis et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text