Selective Backpropagation for Efficient Few-Shot Class-Incremental Learning

Eeham Khan; Abdulmoumen Al-Atrash; Ali Ayub

cs.LGOct 2, 2026

Selective Backpropagation for Efficient Few-Shot Class-Incremental Learning

Authors: Eeham Khan, Abdulmoumen Al-Atrash, Ali Ayub

Organizations: PaInt Lab, Concordia University Montreal, QC, Canada

Abstract

Few-Shot Class-Incremental Learning (FSCIL) requires models to continuously learn new classes from limited samples while retaining prior knowledge, under strict constraints on compute and memory. Existing approaches lie along a difficult trade-off: simple fine-tuning is computationally efficient but suffers from catastrophic forgetting, replay-based methods mitigate forgetting at the cost of substantial compute and memory, and exemplar-free methods often reduce forgetting by freezing most of the backbone, improving efficiency at the expense of adaptability. We propose Selective Backpropagation (SBP), a deterministic parameter budgeting framework that bridges this gap. SBP restricts gradient updates to a pre-allocated subset of network parameters, freezing past knowledge and preserving unbiased capacity for future learning, enabling rapid adaptation without costly mask optimization. We show that SBP achieves strong performance across standard FSCIL benchmarks while requiring training time close to that of naive fine-tuning and substantially lower training time than prior SOTA methods. Crucially, our experiments expose a limitation of standard FSCIL evaluation: performance on short, distribution-consistent benchmarks does not necessarily predict behavior under distribution shift or over substantially longer learning horizons. We therefore evaluate FSCIL methods in cross-domain settings and over an 80-session ImageNet-1K stream. SBP remains strong across these regimes while maintaining low training cost, providing a favorable stability-plasticity-efficiency trade-off. Our code is available at https://github.com/PaInt-Lab/sbp-main-public.

Figures & tables

Figure 1: Selective Backpropagation (SBP) overview: parameter partitioning and masked gradient flow across incremental sessions.

Method	Sessions (Accuracy %)									Efficiency
	1	2	3	4	5	6	7	8	9	Avg	Time (m)	vs Ours
Entropy-Reg Liu et al. (2022)	74.37	71.31	68.93	66.61	64.43	62.30	60.29	58.40	56.60	64.80	486	$34.71\times$
ALICE Peng et al. (2022)	79.00	70.50	67.10	63.40	61.20	59.20	58.10	56.30	54.10	63.21	196	$14.00\times$
FACT Zhou et al. (2022a)	80.13	75.09	70.51	66.43	62.55	59.33	56.87	54.42	51.95	64.14	40	$2.86\times$
WaRP Kim et al. (2023)	80.31	75.86	71.87	67.58	64.39	61.34	59.15	57.10	54.74	65.82	20	$1.43\times$
NC-FSCIL Yang et al. (2023)	79.57	59.23	61.03	61.76	58.79	55.28	53.74	51.96	49.43	58.98	85	$6.07\times$

Table 1: CIFAR-100 FSCIL (5-way 5-shot) classification accuracy (%) with ResNet-18 using the TOPIC Cheraghian et al. (2021) class split. Results are over five runs with different random seeds. Best (per column) is in bold , second best is underlined . For efficiency, lower is better.

Method	Sessions (Accuracy %)									Efficiency
	1	2	3	4	5	6	7	8	9	Avg	Time (m)	vs Ours
Entropy-Reg Liu et al. (2022)	72.49	66.93	62.21	58.24	54.74	51.59	48.86	46.42	44.43	56.21	573	$19.10\times$
ALICE Peng et al. (2022)	80.60	70.60	67.40	64.50	62.50	60.00	57.80	56.80	55.70	63.99	314	$10.47\times$
FACT Zhou et al. (2022a)	70.45	64.73	60.58	57.44	54.32	51.60	49.00	47.00	45.46	55.62	41	$1.37\times$
WaRP Kim et al. (2023)	68.63	64.43	60.44	57.82	55.42	52.64	50.02	48.19	46.89	56.05	27	$0.90\times$
NC-FSCIL Yang et al. (2023)	76.45	68.12	63.50	59.28	57.10	54.22	51.80	49.65	47.90	58.67	883	$29.43\times$

Table 2: MiniImageNet FSCIL (5-way 5-shot) classification accuracy (%) and training efficiency with ResNet-18. Results are over five runs with different random seeds. Best (per column) is in bold , second best is underlined . For efficiency, lower is better.

Method	Sessions (Accuracy %)											Avg (%)
	1	2	3	4	5	6	7	8	9	10	11
ALICE Peng et al. (2022)	65.56	56.05	51.73	47.68	44.96	41.81	39.37	37.19	35.26	33.45	31.88	44.09
FACT Zhou et al. (2022a)	79.20	62.46	51.83	44.39	39.07	35.06	31.38	28.70	26.13	24.41	22.57	40.47
WaRP Kim et al. (2023)	69.41	64.13	59.33	54.81	51.37	48.23	45.54	43.06	41.09	38.39	36.12	50.13
NC-FSCIL Yang et al. (2023)	64.35	54.27	50.17	44.15	39.47	35.64	32.46	29.82	27.37	25.47	23.81	38.82
OrCo Ahmed et al. (2024)	58.05	43.95	39.89	34.38	30.35	29.62	28.33	26.50	24.25	23.24	22.07	32.78

Table 3: Cross-Domain FSCIL (MiniImageNet

\rightarrow

CUB-200) classification accuracy (%) with an initial session of 60 base classes and 10 incremental sessions of 20-way 5-shot using a ResNet-18 backbone. Best (per column) is in bold , second best is underlined .

Method	Sessions (Accuracy %)
	0	1	5	10	20	30	40	50	60	70	80
ADBS Li et al. (2025a)	61.50	57.10	47.20	39.20	27.40	22.20	18.80	17.20	16.00	14.90	14.50
OrCo Ahmed et al. (2024)	27.00	24.80	24.90	25.10	24.30	23.20	21.90	21.00	20.00	19.30	18.40
WaRP Kim et al. (2023)	63.20	61.90	60.40	58.60	54.90	51.40	47.80	42.10	33.30	20.00	12.90
Ours	72.45	70.09	64.99	61.85	54.26	50.96	47.93	45.19	43.01	40.83	39.08

Table 4: Long-horizon FSCIL on ImageNet-1K. We report classification accuracy (%) at selected sessions throughout the incremental sequence. Best (per column) is in bold , second best is underlined .

	Session Accuracy (%)									Metrics
Ablation	1	2	3	4	5	6	7	8	9	Avg (%)	$\Delta$
Full	82.70	78.40	74.40	70.90	67.50	64.90	62.00	59.90	57.90	68.73	—
SBP Only	82.85	76.25	70.81	66.20	61.90	58.20	54.36	51.47	48.85	63.43	$-$ 5.30
w/o Imprinting	82.70	75.75	69.99	65.01	60.70	57.18	53.76	50.77	48.24	62.63	$-$ 6.10
w/o SBP	84.10	13.90	10.00	5.60	7.00	6.20	4.00	4.00	4.30	15.46	$-$ 53.27
w/o Free-Init	82.60	1.88	1.66	1.69	1.85	1.36	1.56	1.43	1.56	10.62	$-$ 58.11

Table 5: Ablation study on CIFAR-100 FSCIL, 5-way 5-shot task. We report per-session classification accuracy (%), average incremental accuracy, and the change in average accuracy relative to the full model (

\Delta

). Best (per column) is in bold , second best is underlined .

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Backbone	Sessions (time in minutes)									Total
		1	2	3	4	5	6	7	8	9	(min)
Entropy-Reg Liu et al. (2022)	Tuned	56	52	52	53	53	54	55	55	56	486
ALICE Peng et al. (2022)	Frozen	188	1	1	1	1	1	1	1	1	196
FACT Zhou et al. (2022a)	Frozen	39	0.03	0.03	0.03	0.03	0.03	0.03	0.03	0.03	40
WaRP Kim et al. (2023)	Tuned	19	0.08	0.05	0.05	0.05	0.05	0.05	0.05	0.06	20
NC-FSCIL Yang et al. (2023)	Frozen	84	0.03	0.04	0.06	0.06	0.07	0.08	0.08	0.08	85

Appendix

Table 6: CIFAR-100 FSCIL (5-way 5-shot) per-session runtime (minutes) with ResNet-18 using the TOPIC class split. The “Backbone Update” column indicates whether the backbone is tuned during incremental sessions or frozen after base training. Session 1 reports the base session (60 classes) runtime. Lower runtime is better; best and second-best values in each runtime column are shown in bold and underlined , respectively. Results are averaged over five runs with different random seeds.

Method	Sessions (Accuracy %)											Avg (%)
	1	2	3	4	5	6	7	8	9	10	11
ALICE Peng et al. (2022)	67.63	58.42	42.51	37.58	55.33	36.24	55.90	49.83	48.47	65.97	58.00	52.35
FACT Zhou et al. (2022a)	65.02	62.23	58.92	55.14	54.41	51.97	51.50	49.02	49.25	48.53	47.85	53.99
WaRP Kim et al. (2023)	77.65	74.80	71.60	67.44	64.92	61.88	60.94	58.75	55.86	53.66	49.91	63.40
NC-FSCIL Yang et al. (2023)	78.42	71.08	67.41	63.49	59.91	57.61	55.29	54.39	51.23	49.13	48.02	59.63
OrCo Ahmed et al. (2024)	73.01	65.70	63.64	61.94	60.45	59.22	58.41	57.59	55.84	56.44	56.11	60.76

Appendix

Table 7: CUB-200-2011 FSCIL (10-way 5-shot) classification accuracy (%) with ResNet-18. Best (per column) is in bold , second best is underlined . Results are averaged over five runs with different random seeds.

Backbone	Sessions (Accuracy %)									Avg (%)
	1	2	3	4	5	6	7	8	9
Without pretraining
CNNx4	61.92	58.75	56.13	53.72	50.93	48.62	45.62	43.95	42.03	51.30
CNNx8	65.30	61.91	59.14	56.52	54.10	51.78	49.21	47.58	45.89	54.60
CNNx16	68.05	64.43	61.76	58.93	56.13	53.82	51.18	49.44	47.50	56.80
ResNet-18	82.50	77.22	72.69	70.55	67.76	64.40	62.76	60.31	58.52	68.52

Appendix

Table 8: Architecture study of our method on CIFAR-100 FSCIL (5-way 5-shot). We report classification accuracy (%) using different backbone networks, including CNNx4, CNNx8, CNNx16, ResNet-18, ResNet-50, and ResNeXt-29. Best (per column) is in bold , second best is underlined .

Method	Sessions (Accuracy %)											Avg (%)
	0	1	2	3	4	5	6	7	8	9	10
ALICE Peng et al. (2022)	76.50	71.60	68.90	66.40	63.70	61.70	59.50	57.60	55.80	54.30	53.00	62.64
OrCo Ahmed et al. (2024)	76.50	63.60	61.40	59.20	58.00	57.00	55.40	54.50	53.10	51.80	50.30	58.25
ADBS Li et al. (2025a)	82.10	68.30	61.60	50.70	44.20	41.60	37.80	33.50	30.20	28.90	26.60	45.95
WaRP Kim et al. (2023)	69.41	64.41	63.15	61.86	60.06	58.70	57.68	57.22	56.11	55.19	54.39	59.83
Ours	77.33	70.34	67.88	66.70	65.98	64.27	62.86	60.47	59.76	59.21	58.77	64.87

Appendix

Table 9: Cross-Domain FSCIL (MiniImageNet

\rightarrow

Caltech-101) classification accuracy (%) with a ResNet-18 backbone. Best (per column) is in bold , second best is underlined .

ResNet-18 backbone
Method	Sessions (Accuracy %)
	0	1	5	10	20	30	40	50	60	70	80
ADBS Li et al. (2025a)	61.50	57.10	47.20	39.20	27.40	22.20	18.80	17.20	16.00	14.90	14.50
OrCo Ahmed et al. (2024)	27.00	24.80	24.90	25.10	24.30	23.20	21.90	21.00	20.00	19.30	18.40
WaRP Kim et al. (2023)	63.20	61.90	60.40	58.60	54.90	51.40	47.80	42.10	33.30	20.00	12.90
SBP (ResNet-18)	72.45	70.09	64.99	61.85	54.26	50.96	47.93	45.19	43.01	40.83	39.08

Appendix

Table 10: Long-horizon FSCIL on ImageNet-1K. Session 0 denotes the base session, followed by 80 incremental sessions. ADBS, OrCo, WaRP, and SBP (ResNet-18) use the same ResNet-18 backbone and are directly comparable. SBP (ViT-B/16) is reported separately as an additional backbone experiment and is not included in the best/second-best highlighting of the ResNet-18 comparison.

Budget	Sessions (Accuracy %)									Avg (%)
	1	2	3	4	5	6	7	8	9
50%	74.87	71.20	68.09	65.32	62.14	59.28	56.14	53.68	52.14	62.54
75%	81.53	76.98	73.26	69.55	66.09	63.69	60.21	57.84	55.87	67.22
80%	81.93	77.62	74.01	70.47	66.81	64.39	61.37	59.03	57.01	68.07
85%	82.65	77.98	74.31	70.89	67.35	64.36	60.77	58.47	57.84	68.29
90%	82.32	77.95	74.17	70.75	67.06	64.73	61.73	59.37	56.89	68.33

Appendix

Table 11: Impact of base session capacity budget on CIFAR-100 FSCIL performance. We vary the fraction of network capacity allocated to the base session and report per-session classification accuracy (%) and average accuracy over all sessions. Best (per column) is in bold , second best is underlined .

Panel of Figure 2 — Figure 2: Evolution of base-class features with and without SBP’s free-parameter reinitialization. Each point represents a sample, with color indicating its class. Without reinitialization, base classes are initially separated after Session 1, but collapse after Session 2 once the implicitly co-adapted free parameters are reassigned and updated. With epoch-wise reinitialization, the feature clusters remain compact and well separated after Session 2, indicating that the learned representation is invariant to changes in the free subspace.

Class-Incremental Learning

Class-incremental learning studies how models can add new categories over time while retaining the ability to recognize previously learned classes.

See all 159 papers

Decoupled and Distilled: Task-Adaptive LoRA-Teachers with Ensemble Knowledge Transfer for Few-Shot Class-Incremental Learning
Sep 30, 2026Hongwei Zhao, Rui Liu, Yansong Liu +2Class-Incremental Learning Low-Rank Adaptation
School of Computer Science and Engineering, Beihang University · School of Computer Science, Beijing University of Posts and Telecommunications
Learning to Curate What You Generate for Generalizable Few-Shot Class-Incremental Learning
Oct 4, 2026Junhui Yin, Yuchen Yang, Yilin Yin +6Class-Incremental Learning Data Selection
Beijing University of Posts and Telecommunications, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Singapore Management University, Singapore, Singapore
Alleviating Regional Shortcuts for Few-Shot Class-Incremental Learning
Jul 24, 2026Haichen Zhou, Yazhe Lyu, Yixiong Zou +2Class-Incremental Learning Incremental Learning
School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan 430074, China

Sep 30, 2026J/K move · Enter open · S save

Decoupled and Distilled: Task-Adaptive LoRA-Teachers with Ensemble Knowledge Transfer for Few-Shot Class-Incremental Learning

Hongwei Zhao, Rui Liu, Yansong Liu, Zhiyuan Zou+1

School of Computer Science and Engineering, Beihang University · School of Computer Science, Beijing University of Posts and Telecommunications

Class-Incremental Learning Low-Rank Adaptation Multi-Teacher Knowledge Distillation Few-Shot Image Classification Knowledge Distillation

Open paper AskPDF arXiv

Few-Shot Class-Incremental Learning (FSCIL) addresses the challenge of learning new classes from very limited samples while retaining knowledge of previously learned ones. Although parameter-efficient fine-tuning methods with pre-trained models show promise for class-incremental learning, strict gradient-based constraints can be unreliable under severe data scarcity, while multi-expert approaches can impose substantial inference-time costs. We propose TALON (Task-Adaptive LoRA-Teachers with Ensemble Knowledge Transfer), an inference-efficient FSCIL framework. TALON dynamically allocates an independent LoRA-Teacher to each incremental task for task-specific representation learning, then distills multiple frozen teachers into a unified LoRA-Student through Ensemble Knowledge Transfer, eliminating runtime module selection or generation. A semantic-guided distillation strategy weights teacher contributions by feature-space similarity to mitigate catastrophic forgetting and overfitting. Across three class-order runs, TALON achieves comparable or better mean average accuracy across four FSCIL benchmarks, obtaining 86.68 +/- 1.22% on CUB200, 90.39 +/- 0.27% on CIFAR100, 78.38 +/- 0.94% on ImageNet-R, and 96.34 +/- 0.33% on miniImageNet. TALON uses up to 33x fewer deployment parameters and reduces average inference time per task to 26.7 s, a 41.70% reduction relative to ASP.

Abstract

Class-Incremental Learning

Explore similar work

Decoupled and Distilled: Task-Adaptive LoRA-Teachers with Ensemble Knowledge Transfer for Few-Shot Class-Incremental Learning

Learning to Curate What You Generate for Generalizable Few-Shot Class-Incremental Learning

Alleviating Regional Shortcuts for Few-Shot Class-Incremental Learning