Grow, Assess, Compress: Adaptive Backbone Scaling for Memory-Efficient Class Incremental Learning

Adrian Garcia-Castañeda; Jon Irureta; Jon Imaz; Aizea Lojo

cs.LGMar 9, 2026

Grow, Assess, Compress: Adaptive Backbone Scaling for Memory-Efficient Class Incremental Learning

Authors: Adrian Garcia-Castañeda, Jon Irureta, Jon Imaz, Aizea Lojo

Organizations: Ikerlan Technology Research Center, Arrasate, Spain · Goi Eskola Politeknikoa, Mondragon Unibertsitatea, Arrasate, Spain · University of the Basque Country UPV/EHU, Spain

Abstract

Class Incremental Learning (CIL) poses a fundamental challenge: maintaining a balance between the plasticity required to learn new tasks and the stability needed to prevent catastrophic forgetting. While expansion-based methods effectively mitigate forgetting by adding task-specific parameters, they suffer from uncontrolled architectural growth and memory overhead. In this paper, we propose a novel dynamic scaling framework that adaptively manages model capacity through a cyclic "GRow, Assess, ComprEss" (GRACE) strategy. Crucially, we supplement backbone expansion with a novel saturation assessment phase that evaluates the utilization of the model's capacity. This assessment allows the framework to make informed decisions to either expand the architecture or compress the backbones into a streamlined representation, preventing parameter explosion. Experimental results demonstrate that our approach achieves state-of-the-art performance across multiple CIL benchmarks, while reducing memory footprint by up to a 73% compared to purely expansionist models.

Figures & tables

Panel of Figure 1 — Figure 1: Mean average accuracy (top left) and parameter count (top right) on CIFAR-100 across different split protocols. The bottom panel illustrates the expansion dynamics for the Base 50 Inc 2 setting. Unlike baseline methods that expand arbitrarily each task or remain static, our proposed framework selectively grows the network only when necessary, obtaining state-of-the-art performance with a significantly lower parameter count.

Figure 2: Overview of the proposed GRACE framework (best viewed in colour). The incorporation of each novel task follows a three-stage pipeline: (1) Grow , the architecture expansion phase (Section 3.3 ); (2) Assess , the capacity evaluation phase (Section 3.4 ); and (3) Compress , the model consolidation phase (Section 3.5 ).

Figure 3: Model consolidation phase of the proposed GRACE framework. The expanded model (

f_{exp}

) acts as a teacher to guide the training of the compressed student model (

f_{com}

). Knowledge transfer is facilitated through a multi-objective loss function: logit-level distillation (

\mathcal{L}_{kd}

), feature-level alignment (

\mathcal{L}_{feat}

), and cross-entropy loss (

\mathcal{L}_{ce}

) to maintain classification performance.

Method	Base 0 Inc 5			Base 0 Inc 10			Base 50 Inc 2			Base 50 Inc 10
Method	#P	Last	Avg	#P	Last	Avg	#P	Last	Avg	#P	Last	Avg
Finetune	0.46	$4.94$	$17.39$	0.46	$8.86$	$26.11$	0.46	$1.94$	$5.79$	0.46	$8.88$	$22.51$
Replay	0.46	$37.25$	$57.00$	0.46	$40.05$	$58.65$	0.46	$40.12$	$48.93$	0.46	$42.11$	$52.09$
EWC	0.46	$4.90$	$18.73$	0.46	$12.41$	$30.82$	0.46	$8.50$	$18.37$	0.46	$14.41$	$29.32$
iCaRL	0.46	$45.34$	$63.08$	0.46	$48.64$	$65.06$	0.46	$43.45$	$53.35$	0.46	$51.90$	$61.04$
DER	9.20	56.89	69.31	4.60	60.47	71.23	11.96	58.95	66.09	2.76	62.17	68.68

Table 1: Experimental results for CIFAR-100 dataset. The best and second-best accuracy results per scenario are highlighted in bold and underlined , respectively.

Method	Base 0 Inc 5			Base 0 Inc 10			Base 50 Inc 10
Method	#P	Last	Avg	#P	Last	Avg	#P	Last	Avg
Finetune	11.17	$4.66$	$16.82$	11.17	$9.28$	$26.28$	11.17	$9.58$	$24.53$
Replay	11.17	$36.34$	$55.62$	11.17	$42.64$	$60.26$	11.17	$41.46$	$51.86$
EWC	11.17	$5.60$	$18.83$	11.17	$10.86$	$28.06$	11.17	$11.68$	$26.35$
iCaRL	11.17	$44.48$	$61.86$	11.17	$50.46$	$67.23$	11.17	$53.96$	$63.42$
DER	223.4	6 4.62	73.66	111.7	67.00	76.16	67.02	7 3.34	79.59

Table 2: Experimental results for ImageNet-100 dataset. The best and second-best accuracy results per scenario are highlighted in bold and underlined , respectively. Results for BEEF are omitted for the Base 0 Inc 5 and Base 50 Inc 10 settings, as the model’s memory requirements exceeded our 48GB GPU capacity.

Method	CIFAR-100
	Base 0 Inc 5			Base 0 Inc 10
	M	Last	Avg	M	Last	Avg
Replay	$13\,466$	$60.83$	$73.66$	$7431$	$56.37$	$70.26$
iCaRL	$13\,466$	$61.40$	$74.02$	$7431$	$59.08$	$71.41$
DER	$2000$	$56.89$	$69.31$	$2000$	$60.47$	$71.23$
FOSTER	$13\,466$	6 3.28	7 4.94	$7431$	$60.26$	7 2.56

Table 3: Memory-aligned comparison results. Comparison on CIFAR-100 and ImageNet-100 under identical memory constraints. Differences in parameter count are reallocated to the rehearsal memory size, M, to ensure an equitable evaluation.

Panel of Figure 4 — Figure 4: GRACE’s expansion dynamics (left) and accuracy change (right) on CIFAR-100 Base 0 Inc 10 with different threshold and decay configurations.

Method	Components			Last	Avg
Method	Init.	Feat KD	Logit KD	Last	Avg
DER (Base)	–	–	–	60.47	71.23
GRACE	✗	✗	✓	38.41	58.23
	✗	✓	✗	45.36	60.03
	✓	✗	✗	54.11	68.69
	✗	✓	✓	47.64	61.68

Table 4: Ablation results on GRACE compression. Comparison of different compression components in CIFAR-100 under the Base 0 Inc 10 setting.

Method	CIFAR-100 Base 0
	Inc 5		Inc 10
	Last	Avg	Last	Avg
Avg. Baseline	$56.99$	$69.50$	$60.02$	$70.64$
Imp.-Aw. (Ours)	57.39	69.81	60.80	71.44
$\Delta$	+0.40	+0.31	+0.78	+0.80
Method	CIFAR-100 Base 50

Table 5: Impact of Importance-Aware Student Initialization. Evaluation across four CIFAR-100 benchmarks ( Base 0 Inc 5/10 and Base 50 Inc 2/10 ). Results contrast standard average initialization with our proposed importance-aware technique.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Comparison of different saturation measures over CIFAR-100 Base 0 Inc 5 , with fixed expansions in tasks 6 and 13.

Panel of Figure 6 — Figure 6: Sensitivity analysis of threshold and threshold decay values on CIFAR-100. The plots compare performance across different protocol settings, highlighting the Pareto frontier for hyperparameter selection. Note that some points may be overlapped, meaning that runs with different threshold values produced equal compression decisions.

Method	Base 0 Inc 10			Base 50 Inc 10
Method	#P	Last	Avg	#P	Last	Avg
DER	$4.60$	6 8.77	77.12	$2.76$	$69.07$	$74.07$
TagFex	$4.60$	68.86	7 7.05	$2.76$	6 9.28	74.42
CREATE	$0.87$	$55.64$	$69.88$	$0.87$	$57.86$	$67.74$
GRACE	$2.30$	$67.98$	$77.00$	$1.84$	69.42	7 4.10

Appendix

Table 6: CIFAR-100 results with AutoAugment. Comparative analysis for Base 0 Inc 10 and Base 50 Inc 10 incremental settings. The best and second-best accuracy results per scenario are highlighted in bold and underlined , respectively.

Storage-Efficient Continual Learning

See all 11 papers

Width Expansion as a Method for Class Incremental Learning
Sep 29, 2026A. L. S. Conde, Y. Elkhatib, C. M. RanieriClass-Incremental Learning Dynamic Neural Networks
Institute of Geosciences and Exact Sciences, São Paulo State University (UNESP), Av. 24-A, 1515, Rio Claro, 13506-692, SP, Brazil · School of Computing Science, University of Glasgow, Glasgow, G12 8QQ, Scotland, United Kingdom
Semantic-Guided Dynamic Sparsification for Pre-Trained Model-based Class-Incremental Learning
Jan 29, 2026Ruiqi Liu, Boyu Diao, Zijia An +3Class-Incremental Learning Activation Sparsity
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China
Miles: Metric Learning with Expandable Subspace for Pre-Trained Model-Based Class-Incremental Learning
Jul 20, 2026Kai Jiang, Zisong Lin, Hongyuan Zhang +2Class-Incremental Learning Metric Learning
National Key Laboratory of Radar Signal Processing, Xidian University, Xi’an 710071, China · The University of Hong Kong, Hong Kong SAR, China · Institute of Artificial Intelligence (TeleAI) of China Telecom, China

Sep 29, 2026J/K move · Enter open · S save

Width Expansion as a Method for Class Incremental Learning

A. L. S. Conde, Y. Elkhatib, C. M. Ranieri

Institute of Geosciences and Exact Sciences, São Paulo State University (UNESP), Av. 24-A, 1515, Rio Claro, 13506-692, SP, Brazil · School of Computing Science, University of Glasgow, Glasgow, G12 8QQ, Scotland, United Kingdom

Class-Incremental Learning Dynamic Neural Networks Incremental Learning

Open paper AskPDF arXiv

Class Incremental Learning (Class-IL) requires models to learn new classes over time while preserving previously acquired knowledge without access to past data or task identity. This setting intensifies the stability-plasticity dilemma and makes catastrophic forgetting a central challenge. Existing approaches include regularization, knowledge distillation, replay, and architectural expansion. However, many expansion methods rely on explicit task identifiers or predefined growth strategies, limiting their applicability when task boundaries are unavailable at inference time. This work proposes a dynamic width expansion method that increases the number of neurons within existing layers according to a normalized loss criterion, without requiring task-specific information. An attention mechanism with persistent key-value memory is also incorporated to stabilize feature representations and reduce interference between previously learned and newly introduced classes. The approach is evaluated on Split MNIST and Split CIFAR-100 under the standard Class-IL protocol. Experiments compare fixed-capacity and dynamically expanding architectures, both with and without attention, combined with established continual learning methods including EWC, LwF, and A-GEM. Results show that progressive width expansion consistently improves performance over fixed architectures, particularly when combined with functional methods and A-GEM. The combination of width expansion and attention provides the most consistent gains. Overall, dynamic width expansion based on representational demand provides an effective and flexible strategy for Class-IL, although uncontrolled growth may increase overfitting and computational cost.

Abstract

Storage-Efficient Continual Learning

Explore similar work

Width Expansion as a Method for Class Incremental Learning

Semantic-Guided Dynamic Sparsification for Pre-Trained Model-based Class-Incremental Learning

Miles: Metric Learning with Expandable Subspace for Pre-Trained Model-Based Class-Incremental Learning