We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset with imbalance factor IF drives model parameters into an O(1/IF) neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we establish an objective-level reformulation of LTR as a sequential learning problem and prove that the CL objective upper-bounds the balanced LTR loss up to a constant determined by the residual Head loss. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.
Figures & tables
Figure 1: Overview of our framework: (left) theoretical analysis of convergence under the LTR setup, showing how imbalance drives convergence toward head-class solutions; (middle) our proposed CLTR reformulating LTR as a CL problem; and (right) the conditional optimization guarantee establishing that the CL objective upper-bounds the balanced LTR loss up to an additive constant.
Figure 2: Left: Illustration of the Head loss landscape ( LH ) for convex setting (Assumption 3.2 ); the unique minimizer ( θH⋆ ) is marked in orange. Middle: Illustration of the Head loss landscape for feedforward deep network, KL((1/2)) setting (Assumption 3.4 ); multiple local minimizers ( θH,k⋆ ) are shown in orange. In both panels, the red region denotes the provable neighborhood that contains the parameter ( θ⋆ ) obtained when training on the full long-tailed dataset; its radius scales as ( O(1/IF) ) by Theorems 3.3 and 3.5 . (Top: surface; bottom: contour projection.). Right: Overview of learning under the LTR scenario and our proposed CLTR approach (symbols described in the text).
Figure 3: Empirical validation for Theorems 3.3 and 3.5 . Top: log-reg on MNIST-LT with varying imbalance factors and L2 strengths μ . Bottom: the same analysis for ResNet-18 on CIFAR-100, with and without weight decay.
Model
IF
100
50
10
Baseline [ 16 ]
38.3
43.9
55.7
Baseline + CB [ 16 ]
39.6
45.3
58.0
Focal loss [ 41 ]
38.4
44.3
55.8
Focal+CB [ 16 ]
39.6
45.2
58.0
τ -norm [ 21 ]
47.7
52.5
63.8
Table 1: LTR benchmarks for (left) CIFAR100-LT, (middle) CIFAR10-LT, and (right) ImageNet-LT. Bold and underlined values denote the best and second-best results in each column.
Method
IF
100
50
10
Zero-shot CLIP
64.5
64.5
64.5
Linear probe
60.0
67.2
75.6
Full fine-tuning
73.4
74.8
81.0
Prompt learning
60.4
63.1
66.9
Naive sequential
48.0
51.2
53.1
Table 2: Foundation-model adaptation on CIFAR100-LT. Bold denotes the best result without additional semantic supervision. Gray rows are contextual, non-protocol-identical comparisons.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Algorithm
LR
Opt.
Momentum
LR Scheduler
CL Loss Weight
Epochs
Steps
LwF
0.001
SGD
0.9
–
0.01
5
2
EWC
0.01
SGD
0.9
–
10
90
2
Modified EWC
0.01
SGD
0.9
–
1000
90
2
GPM
0.001
SGD
0
Cosine Anneal LR
–
100
2
SGP
0.001
SGD
0.9
Cosine Anneal LR
–
100
2
FOSTER (phase 1)
0.1
SGD
0.9
Cosine Anneal LR
–
170
4
Appendix
Table A1 : Implementation details of the considered algorithms for the LTR benchmark.
Method
Backbone
Trainable params
Opt.
LR
Epochs
Tasks
Zero-shot CLIP
ViT-B/32 (frozen)
0
–
–
–
–
Linear probe
ViT-B/32 (frozen)
head only
AdamW
10−3
50
1
Full fine-tuning
ViT-B/32
all
AdamW
3×10−5
20
1
Prompt learning
ViT-B/32 (frozen)
visual prompts
SGD
40
10
1
Naive sequential
ViT-B/32
all
AdamW
3×10−5
20
2
DualPrompt
ViT-B/32 (LN adapted)
G/E prompts + LN + cosine head
AdamW
3×10−3
40–60
2
Appendix
Table A2 : Implementation details for foundation-model adaptation experiments on CIFAR100-LT.
Figure A1 : Class cardinality of (a) MNIST-LT, (b) CIFAR100-LT, (c) CIFAR10-LT, (d) ImageNet-LT, and (e) Caltech256.
Method
ImageNet-LT
CIFAR100-LT
IF=100
IF=50
IF=10
CLTR (LwF)
47.6
45.1
49.3
58.7
CLTR (EWC)
48.9
44.4
50.3
58.8
CLTR (Modified EWC)
49.1
45.9
51.0
60.7
CLTR (GPM)
51.7
48.3
54.7
64.7
CLTR (FOSTER)
52.7
48.7
54.4
63.6
Appendix
Table A3 : Performance comparison on CIFAR100-LT and ImageNet-LT. The blue rows show CLTR variants. The gray block reports methods with additional bells and whistles and is included for context, since these methods may not follow the same protocol.
Method
Tasks
5
10
EEIL [ 92 ]
38.5
37.5
EEIL+2sLWS [ 88 ]
39.0
37.6
LUCIR [ 90 ]
42.7
42.2
PODNET [ 91 ]
44.1
44.0
PODNET+2sLWS [ 88 ]
44.4
44.4
Appendix
Table A4 : The performance of CLTR on Ordered LT-CIL Benchmark for CIFAR100-LT
Method
Backbone
Inc.V4
Res.101
L2−FE [ 93 ]
84.1
85.3
L2 [ 93 ]
85.8
87.2
L2−SP [ 93 ]
85.3
87.2
DELTA [ 93 ]
86.8
88.7
GBN [ 94 ]
-
86.9
Appendix
Table A5 : The performance of CLTR on Caltech256.
Training order
Top-1 accuracy
Tail → Head
36.0
Random
32.3
Head → Tail (CLTR order)
50.7
Appendix
Table A6 : Effect of task order for CLTR(SGP) on CIFAR100-LT with IF=100 . Head → Tail follows the cardinality-sorted ordering motivated by our theory and achieves the best performance.
Method
Min. margin
Satisfied
Naive sequential
−12.15
0/33
CLTR (LwF)
+2.25
33/33
CLTR (EWC)
+0.07
33/33
Appendix
Table A7 : Empirical verification of the upper-bound inequality on CIFAR100-LT ( IF=100 ) across 33 measurements per method.
Figure A2 : CLTR performance for various (left) number of incremental steps and (right) replay memory ( M ) size.
Head:Tail split
Local IFs
Top-1 (%)
20:80
[2.43, 39.40]
46.0
50:50
[9.8, 9.6]
49.7
80:20
[41.67, 2.40]
44.7
Appendix
Table A8 : Sensitivity of CLTR (SGP) to Head–Tail class boundaries on CIFAR100-LT ( IF=100 ) with two partitions.
Method
Time (s)
Rel.
FLOPs (T)
Rel.
Mem. (MB)
LTR (WD)
205.5±1.7
0.986
227.0
1.000
334
LTR (WD+MaxNorm)
208.3±7.5
1.000
227.0
1.000
334
CLTR (no CL term)
238.6±1.9
1.145
252.1
1.111
334
CLTR (LwF)
238.2±1.4
1.143
264.4
1.164
374
CLTR (EWC)
260.4±13.0
1.250
254.3
1.120
340
CLTR (SGP)
280.9±2.7
1.349
252.5
1.112
343
Appendix
Table A9 : Computational overhead on CIFAR100-LT using ResNet-32 and a single RTX 3090 GPU. Wall-clock time is averaged over three runs. Relative values are normalized to WD+MaxNorm.
Figure A3 : Per-class weight norms of the baseline, SGP, and WD.
Figure A4 : The difference in per-class accuracy of CLTR (SGP) and the baseline model.
Method (ViT-B)
IF
10
50
100
Baseline [ 16 ]
66.02
54.78
50.59
BCE [ 16 ]
64.63
50.65
45.50
CB [ 16 ]
66.30
56.18
50.63
CLTR (TPL)
69.74
59.71
55.94
Appendix
Table A10 : Performance of CLTR with a ViT-B backbone on CIFAR100-LT under different imbalance factors.
Method
Head
Tail
Balanced
Balanced training ( IF=1 )
77.42
66.18
71.80
Long-tailed joint training
81.53
33.20
57.37
Naive sequential
0.00
32.78
16.39
CLTR (LwF)
79.21
41.69
60.45
CLTR (GPM)
77.39
44.72
61.06
Appendix
Table A11 : Performance on 20 Newsgroups-LT ( IF=100 ). Head, Tail, and Balanced denote macro-averaged classification recall (%).
Institute of Trustworthy Embodied AI, Fudan University, Shanghai, China · Shanghai Key Laboratory of Multimodal Embodied AI, Shanghai, China · Zhejiang University of Technology, Hangzhou, China +2