Dynamic validation and relevance maintenance of neural correlates under continual learning
| dc.contributor.author | Verbytskyi O. S. | |
| dc.contributor.author | Gaidaienko O. V. | |
| dc.date.accessioned | 2026-06-09T10:17:33Z | |
| dc.date.issued | 2025-09-25 | |
| dc.description | Verbytskyi, O. S. Dynamic validation and relevance maintenance of neural correlates under continual learning / O. S. Verbytskyi, O. V. Gaidaienko // Матеріали ХVІ міжнар. науково-технічна конф. "Інновації в суднобудуванні та океанотехніці". – Миколаїв : НУК. – 2025. – С. 628–631. | |
| dc.description.abstract1 | Deep neural networks trained in static environments suffer from model degradation when faced with evolving data distributions (concept drift). This study proposes a framework for dynamic validation and targeted maintenance of concept-specific neurons, addressing the continual learning challenge. Building on L0-optimization for neuron localization, we introduce a "Concept Adequacy Score" to detect knowledge decay. Upon detecting drift, a localized update mechanism fine-tunes only the minimal neuron set responsible for the concept, avoiding costly full-network retraining. Experiments simulating a continual learning scenario show that our approach maintains high classification accuracy over time, reducing computational costs by an order of magnitude compared to full retraining, thereby enabling efficient and adaptive AI systems. | |
| dc.description.provenance | Submitted by Оксана Гайдаєнко (oksana.gaidaienko@nuos.edu.ua) on 2026-06-08T08:38:53Z workflow start=Step: reviewstep - action:claimaction No. of bitstreams: 1 Вербицкий О_3.pdf: 598988 bytes, checksum: f32b2254240fc184c2c25f95f5db90da (MD5) | en |
| dc.description.provenance | Step: reviewstep - action:reviewaction Approved for entry into archive by Бондар Ольга (olga.bondar@nuos.edu.ua) on 2026-06-09T10:13:18Z (GMT) | en |
| dc.description.provenance | Step: editstep - action:editaction Approved for entry into archive by Бондар Ольга (olga.bondar@nuos.edu.ua) on 2026-06-09T10:16:59Z (GMT) | en |
| dc.description.provenance | Step: finaleditstep - action:finaleditaction Approved for entry into archive by Бондар Ольга (olga.bondar@nuos.edu.ua) on 2026-06-09T10:17:33Z (GMT) | en |
| dc.description.provenance | Made available in DSpace on 2026-06-09T10:17:33Z (GMT). No. of bitstreams: 1 Verbytskyi.pdf: 598988 bytes, checksum: f32b2254240fc184c2c25f95f5db90da (MD5) Previous issue date: 2025-09-25 | en |
| dc.identifier.isbn | 978-966-321-487-0 | |
| dc.identifier.uri | https://eir.nuos.edu.ua/handle/123456789/13023 | |
| dc.language.iso | en | |
| dc.relation.ispartofseries | УДК ; 004.032.26:004.89 | |
| dc.subject | continual learning | |
| dc.subject | concept drift | |
| dc.subject | explainable AI | |
| dc.subject | model degradation | |
| dc.subject | localized finetuning | |
| dc.subject | neural correlates. | |
| dc.title | Dynamic validation and relevance maintenance of neural correlates under continual learning | |
| dc.type | Theses |