Home // International Journal On Advances in Software, volume 19, numbers 1 and 2, 2026 // View article

KAN Control Framework for Continual Learning

Authors:
Evgenii Ostanin
Nebojsa Djosic
Salah Sharieh
Alexander Ferworn

Keywords: Continual Learning; Catastrophic Kolmogorov-Arnold Networks; KAN Control Framework; Spline Freezing; Memory Retention

Abstract:
This study extends the concept of progressively reducing catastrophic forgetting in Kolmogorov-Arnold Networks (KAN) by introducing the KAN Control Framework (KCF). In KCF, the stability-plasticity trade-off is described through a control vector that separates memory-related controls (replay policy, buffer size M, replay intensity ρ) from plasticity-related controls (freezing granularity, freeze ratio k, and importance scoring). Using a two-task Split-MNIST protocol with multi-seed evaluation, replay-only behavior is first mapped over (M,ρ) for standard and balanced replay, where retention improvements are largely driven by rehearsal strength. Progressive spline freezing is then studied as an additional control, and the largest gains are observed in constrained replay scenarios, where memory or replay throughput may be limited, while sufficiently strong replay-only settings can still achieve the lowest forgetting overall. Finally, the way in which the freezing granularity (point-level vs. tensor-level) and scoring rules shape the retention-adaptation trade-off are analyzed and practical guidance is provided for selecting KCF settings under common resource limits. Overall, KCF is intended to make CF configurations in KANs more explicit and reproducible.

Pages: 55 to 66

Copyright: Copyright (c) to authors, 2026. Used with permission.

Publication date: June 30, 2026

Published in: journal

ISSN: 1942-2628