Continual LoRA Merging for Class-Incremental Learning

Continual learning requires a model to acquire new knowledge from a sequence of data without catastrophically forgetting what it has already learned. This thesis considers the rehearsal-free class-incremental setting, in which new classes arrive over successive steps, data from earlier steps cannot be stored, and the model must recognise all classes seen so far without knowing from which step an input comes. Parameter-efficient fine-tuning, and low-rank adaptation (LoRA) in particular, provides a natural way to adapt a frozen pretrained backbone using a small set of trainable parameters. When such adaptation is repeated at every incremental step, however, a further question arises: how should the successive low-rank adaptations be organised across the sequence? To study this question, the thesis compares two strategies. SimpleAvg, used as a strong reference baseline, learns an independent LoRA adaptation at each step and merges them after the sequence. RankExt instead maintains a single persistent LoRA adaptation whose capacity grows over time while previously learned components are retained. Knowledge distillation and orthogonality-based regularisation are investigated as stabilisation mechanisms, and performance is examined through complementary open and restricted evaluation. Experiments have been performed with a frozen CLIP ViT-B/16 backbone and evaluate the proposed comparison across multiple image-classification benchmarks, with CIFAR-100 as the primary controlled benchmark and additional datasets used to assess whether the observed behaviour generalises beyond it. The results indicate that the organisation of low-rank adaptations across incremental steps can substantially influence continual-learning behaviour. The divergence between open and restricted evaluation also suggests that performance degradation should not be interpreted only as loss of previously learned discrimination, since competition across incremental steps can substantially affect the final predictions. The stabilisation mechanisms investigated in this work reduce the observed degradation and improve the behaviour of persistent adaptation. Taken together, these findings emphasise that both adaptation structure and evaluation protocol are central to understanding rehearsal-free continual learning with LoRA.