Regularisation-based class-incremental learning (CIL) methods control the stability-plasticity trade-off through fixed coefficients tuned offline, independently of the continual learning objective and blind to phase-specific variations in data volume, task difficulty, and class similarity. We argue this is a fundamental mismatch and propose REDAR: Reinforcement-Driven Adaptive Regularisation for Class Incremental Learning, which treats regularisation strength selection as a sequential decision problem. At each incremental phase, a lightweight reinforcement learning (RL) agent observes a compact state of current performance metrics and incoming data statistics, and outputs the distillation coefficients for that phase. To respect CIL's constraint on historical data access, the policy is trained entirely on pseudo-CIL tasks constructed from the initial phase, then deployed on the incremental sequence without further adaptation. Experiments on CIFAR-100 and Tiny-ImageNet demonstrate that REDAR reduces forgetting and improves forward transfer over fixed and heuristic baselines, and that policies trained on short pseudo-task horizons transfer effectively to longer incremental runs.
Figures & tables
Figure 1: Overview of REDAR . (a) Reinforcement-driven regularisation is formulated as a finite-horizon Markov Decision Process (MDP), M=(S,A,T,R,γ) , where the state si=(Ai−1,Fi−1,CRi,ISi) summarises the current learning dynamics, the action ai=(αi,βi) specifies the phase-specific regularisation coefficients, and the reward ri is the validation accuracy after CIL training. (b) The CIL model jointly optimises cross-entropy, feature distillation, logit distillation, and margin ranking losses while learning from replay memory and incoming samples. (c) The policy is trained offline on pseudo-CIL tasks using a distributed actor–learner architecture with prioritised experience replay before being deployed on the target incremental sequence.
CIFAR-100 (N=2)
CIFAR-100 (N=5)
CIFAR-100 (N=10)
Method
ACC ↑
FRG ↓
FWT ↑
ACC ↑
FRG ↓
FWT ↑
ACC ↑
FRG ↓
FWT ↑
JT (Upper Bound)
72.87
-
-
72.87
-
-
72.87
-
-
iCaRL [ 23 ]
60.32
28.54
-
51.96
55.29
-
45.40
68.26
-
LUCIR [ 9 ]
66.94
24.55
40.18
63.93
31.92
49.42
61.24
32.13
47.68
WA [ 35 ]
67.2
-
-
64.32
29.85
-
57.77
28.88
-
PodNet [ 5 ]
67.51
30.02
-
61.8
38.99
-
59.22
31.35
-
Table 1: Performance comparison on average accuracy ACC , average forgetting FRG and forward transfer FWT on CIFAR-100.
Tiny-ImageNet (N=5)
Tiny-ImageNet (N=10)
Method
ACC ↑
FRG ↓
FWT ↑
ACC ↑
FRG ↓
FWT ↑
JT (Upper Bound)
68.78
-
-
68.78
-
-
iCaRL [ 23 ]
37.51
31.89
-
42.45
32.77
-
EEIL [ 3 ]
47.4
25.56
-
45.2
25.91
-
LUCIR [ 9 ]
56.76
15.04
56.48
53.41
27.60
68.12
WA [ 35 ]
49.28
40.85
-
38.29
37.24
-
Table 2: Performance comparison on average accuracy ACC , average forgetting FRG and forward transfer FWT on Tiny-ImageNet.
Figure 2: Top: Phase-wise accuracy on CIFAR-100 and Tiny-ImageNet under the N=5 incremental setting. Accuracy is reported as the mean over all classes seen after each phase. Bottom: coefficients ( α , β ) used to train each phase.
Figure 3: Cross-setting transfer evaluation on CIFAR-100. Each cell reports ACC (%) for a policy trained on Ntrain phases and deployed on Ntest phases. Diagonal entries (pink outline) correspond to standard in-distribution evaluation.
Method
ACC ↑
FRG ↓
FWT ↑
Fixed (0.1)
63.93
31.92
49.42
Random Selection
59.84
30.28
43.16
REDAR (Ours)
64.43
28.02
51.77
Table 3: Benefit of varying hyperparameters.
State Representation
Metrics
CR
ACC
FRG
IS
ACC
FRG
FWT
✓
✓
✗
✗
62.48
37.39
45.42
✓
✓
✓
✗
59.97
45.46
39.96
✓
✓
✗
✓
62.49
37.44
48.82
✓
✓
✓
✓
64.43
28.02
51.77
Table 4: Ablation results of the state representation
Class-Incremental Learning (CIL) aims to continuously learn new classes without forgetting previously acquired knowledge. Parameter-efficient fine-tuning with pre-trained models reduces parameter overhead but can suffer from cumulative interference and suboptimal alignment between inference samples and specialized modules. We propose Dynamic LoRA-Experts and Prototype-Ensemble Matching (DLEPEM), a two-stage rehearsal-free framework. DLEPEM allocates a task-specific LoRA-Expert for each incremental task to reduce cross-task interference, then combines frozen pre-trained-model prototypes with task-adaptive LoRA-Expert prototypes for reliable task-level discrimination. Experiments on standard CIL and Few-Shot CIL benchmarks demonstrate strong performance under the evaluated protocols.
Hongwei Zhao, Rui Liu, Yansong Liu
School of Computer Science and Engineering, Beihang University
Class-Incremental Learning (CIL) with pre-trained models (PTMs) aims to sequentially adapt PTMs to new categories without forgetting old knowledge. Built upon PTMs, existing adapter-based methods mainly train models via distinct task-specific adapters, and present a uniform knowledge allocation for each adapter during inference. However, this allocation mechanism ignores the nature of task discrepancy and leads to suboptimal utilization of adapters. Also, under CIL constraint, an allocator is prone to forgetting when tasks evolve. To address these issues, we propose a Non-Forgetting Allocation with Bi-Level Competition (NoFA-BC). NoFA-BC constructs a non-forgetting allocator (NFA) by transforming the allocator training into a recursive least-squares problem and achieves an allocator equivalent to that trained with all data. Based on the NFA, a Bi-Level Competition (BLC) including an intra-task level Winner-Takes-All (WTA) mechanism and inter-task Last-Ones-Fall (LOF) elimination is proposed to provide better allocation of adapter knowledge. WTA extracts the most significant logit within a task to represent the adapter's contribution and LOF suppresses the irrelevant adapters. With BLC, participation ratio of each adapter can be tailored for each input. Moreover, a Stability Enhancement (SE) process is incorporated to further improve the performance of old tasks.
Xiang Tan, Run He, Yawen Cui +6
South China University of Technology · Hong Kong Polytechnic University · Agency for Science, Technology and Research +3
Class-Incremental Learning (CIL) requires models to learn new classes without forgetting old ones. A common method is to freeze a pre-trained model and train a new, lightweight adapter for each task. While this prevents forgetting, it treats the learned knowledge as a simple, unstructured collection and fails to use the relationships between tasks. To this end, we propose the Semantic-guided Adaptive Expert Forest (SAEF), a new method that organizes adapters into a structured hierarchy for better knowledge sharing. SAEF first groups tasks into conceptual clusters based on their semantic relationships. Then, within each cluster, it builds a balanced expert tree by creating new adapters from merging the adapters of similar tasks. At inference time, SAEF finds and activates a set of relevant experts from the forest for any given input. The final prediction is made by combining the outputs of these activated experts, weighted by how confident each expert is. Experiments on several benchmark datasets show that SAEF achieves SOTA performance.
Ruiqi Liu, Boyu Diao, Hangda Liu +3
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China