Evolution-Inspired Sample Competition for Deep Neural Network Optimization

Ying Zheng; Yiyi Zhang; Yi Wang; Lap-Pui Chau

cs.CVApr 14, 2026

Evolution-Inspired Sample Competition for Deep Neural Network Optimization

Authors: Ying Zheng, Yiyi Zhang, Yi Wang, Lap-Pui Chau

Organizations: Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong, China. · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China.

Abstract

Conventional deep network training generally optimizes all samples under a largely uniform learning paradigm, without explicitly modeling the heterogeneous competition among them. Such an oversimplified treatment can lead to several well-known issues, including bias under class imbalance, insufficient learning of hard samples, and the erroneous reinforcement of noisy samples. In this work, we present Natural Selection (NS), a novel evolution-inspired optimization method that explicitly incorporates competitive interactions into deep network training. Unlike conventional sample reweighting strategies that rely mainly on predefined heuristics or static criteria, NS estimates the competitive status of each sample in a group-wise context and uses it to adaptively regulate its training contribution. Specifically, NS first assembles multiple samples into a composite image and rescales it to the original input size for model inference. Based on the resulting predictions, a natural selection score is computed for each sample to characterize its relative competitive variation within the constructed group. These scores are then used to dynamically reweight the sample-wise loss, thereby introducing an explicit competition-driven mechanism into the optimization process. In this way, NS provides a simple yet effective means of moving beyond uniform sample treatment and enables more adaptive and balanced model optimization. Extensive experiments on 12 public datasets across four image classification tasks demonstrate the effectiveness of the proposed method. Moreover, NS is compatible with diverse network architectures and does not depend on task-specific assumptions, indicating its strong generality and practical potential. The code will be made publicly available.

Figures & tables

Fig. 1: (a) Natural selection in ecosystem evolution. (b) Illustration of the natural selection process in training samples. After performing image stitching and scale normalization, the resulting stitched image is fed into the classifier to obtain the softmax probability distribution. The predicted probability associated with each sample’s category is then extracted, and the NS score is subsequently computed through normalization.

p^{(g)}=\mathrm{softmax}\!\left(f_{\theta}(S^{\prime(g)})\right).

Algorithm 1 Natural Selection Score Computation

Fig. 2: The distribution curves of sample weights under different settings. Under the NS-WS (green dashed line) and NS-LF (purple dashed line) strategies, samples with higher-weight values are reinforced during training, while those in lower-weight regions are suppressed. This enables differentiated optimization for samples with varying competitive characteristics.

Method	Basic	w/ NS
Method	Basic	LF-2	WS-2	LF-4	WS-4
ResNet-18 [ 53 ]	68.4	68.4	68.6	68.4	68.9 ( $\uparrow$ 0.5 )
Swin-T [ 57 ]	75.2	75.4	75.5	75.2	75.7 ( $\uparrow$ 0.5 )
VMamba-T [ 58 ]	76.7	76.9	76.9	76.8	77.1 ( $\uparrow$ 0.4 )

TABLE I: Evaluation of NS on ImageNet-1K . “WS” denotes boosting the winner, while “LF” means enhancing the loser. “2/4” indicates grouping two or four samples for competition.

Method	CIFAR-10		CIFAR-100
Method	Basic	w/ NS-WS	Basic	w/ NS-WS
AlexNet [ 51 ]	77.4	78.6 ( $\uparrow$ 1.2 )	43.8	46.0 ( $\uparrow$ 2.2 )
VGG-19 [ 52 ]	93.4	93.8 ( $\uparrow$ 0.4 )	71.7	72.6 ( $\uparrow$ 0.9 )
ResNet-110 [ 53 ]	92.8	93.9 ( $\uparrow$ 1.1 )	70.3	72.1 ( $\uparrow$ 1.8 )
ResNeXt-29 [ 54 ]	95.9	96.2 ( $\uparrow$ 0.3 )	81.4	82.3 ( $\uparrow$ 0.9 )

TABLE II: Evaluation of NS-WS on CIFAR-10/100 with different networks.

Method	Twitter I		Twitter II		Flickr		Instagram		FI		EmoSet
Method	Basic	w/ NS-LF	Basic	w/ NS-LF	Basic	w/ NS-LF	Basic	w/ NS-LF	Basic	w/ NS-LF	Basic	w/ NS-LF
AlexNet [ 51 ]	73.8	75.2	68.5	71.0	83.2	83.8	80.6	81.2	56.4	57.9	67.6	69.2
VGG-16 [ 52 ]	75.3	78.1	74.3	74.9	84.4	84.6	82.5	83.1	59.7	61.4	72.2	73.9
ResNet-50 [ 53 ]	81.3	82.9	71.0	74.3	85.7	86.4	84.6	85.2	65.8	67.2	75.8	76.3
DenseNet-121 [ 55 ]	78.9	79.4	72.8	74.0	85.1	85.6	84.0	85.0	63.2	64.4	74.9	75.2
ViT-B/16 [ 56 ]	74.5	76.4	72.1	72.2	86.2	86.4	84.5	85.3	65.6	67.4	77.3	77.7

TABLE III: Comparison of basic and NS-enhanced models on six emotion recognition datasets.

Method	Ar $\rightarrow$ Cl	Ar $\rightarrow$ Pr	Ar $\rightarrow$ Rw	Cl $\rightarrow$ Ar	Cl $\rightarrow$ Pr	Cl $\rightarrow$ Rw	Pr $\rightarrow$ Ar	Pr $\rightarrow$ Cl	Pr $\rightarrow$ Rw	Rw $\rightarrow$ Ar	Rw $\rightarrow$ Cl	Rw $\rightarrow$ Pr	Avg.
Baseline	58.0	78.8	81.9	69.4	79.7	78.7	68.2	55.0	81.9	73.8	58.7	83.8	72.3
Focal Loss [ 5 ]	57.9	78.8	82.0	69.2	79.8	78.6	68.3	55.4	81.8	73.7	59.9	83.7	72.4
GCE [ 62 ]	57.5	77.7	80.9	68.8	76.5	77.3	66.9	54.5	80.9	73.8	58.7	84.1	71.5
LS [ 63 ]	58.0	78.7	82.0	69.3	79.7	78.6	68.3	55.3	81.9	73.8	59.7	83.9	72.4
NLNL [ 64 ]	56.4	77.5	80.1	67.4	75.6	76.3	65.5	54.8	80.5	73.4	59.0	83.5	70.8
SCE [ 65 ]	54.5	78.3	80.6	63.8	76.7	76.2	65.3	50.4	80.6	70.5	53.7	82.8	69.4

TABLE IV: Comparison of different methods on the Office-Home benchmark.

Method	CIFAR-LT-10				CIFAR-LT-100
Method	200	100	50	10	200	100	50	10
Baseline	64.0	69.5	72.7	86.7	33.6	38.4	42.1	53.8
Focal Loss [ 5 ]	60.2	69.5	73.4	86.9	33.6	37.6	43.6	55.4
GCE [ 62 ]	46.3	60.3	63.8	86.2	18.6	21.8	25.3	36.8
LS [ 63 ]	62.5	70.3	74.8	86.8	32.9	37.8	42.9	56.0
NLNL [ 64 ]	33.0	35.1	39.0	52.6	1.9	1.6	1.6	2.2

TABLE V: Performance comparison on the CIFAR-LT-10/100 datasets.

Group	1 $\times$ 2	2 $\times$ 2	2 $\times$ 4	4 $\times$ 2	4 $\times$ 4
Acc (%)	72.0	72.1	71.7	71.4	71.7

TABLE VI: Evaluation of five different grouped stitching configurations.

$\sigma$	0.0	0.1	0.5	0.8	1.0	1.5	1.8
Acc (%)	67.5	70.9	71.3	72.1	71.7	71.5	71.4

TABLE VII: Performance with different values of

\sigma

$\rho$	0.0	0.1	0.5	0.8	1.0	1.5	1.8
Acc (%)	70.3	71.3	71.6	71.8	72.1	71.9	71.8

TABLE VIII: Performance with different values of

\rho

Dataset	Mean $\pm$ Std	Dataset	Mean $\pm$ Std
ImageNet-1K	73.9 $\pm$ 0.05	Instagram	83.7 $\pm$ 0.21
CIFAR-10	91.8 $\pm$ 0.08	FI	64.2 $\pm$ 0.26
CIFAR-100	70.9 $\pm$ 0.22	EmoSet	75.1 $\pm$ 0.12
Twitter I	77.0 $\pm$ 0.45	Office-Home	73.0 $\pm$ 0.19
Twitter II	73.3 $\pm$ 0.34	CIFAR-LT-10	74.9 $\pm$ 0.43
Flickr	85.5 $\pm$ 0.11	CIFAR-LT-100	43.9 $\pm$ 0.18

TABLE IX: Mean and standard deviation of the performance on the 12 public datasets.

Dataset	ResNet-18	w/ NS-WS
ImageNet-Hard	9.9	10.2

TABLE X: Results on the ImageNet-Hard dataset.

Panel of Figure 14 — Fig. 3: Visualization of NS score distributions by category on the CIFAR-10 and FI datasets. Red points denote training samples, and the blue solid line within each box indicates the median NS score of all samples in that class.

Dataset	TextCNN	w/ NS-LF	BERT-base	w/ NS-LF
IMDB	75.6	76.6	91.8	92.2
AG News	90.3	91.0	94.4	94.7

TABLE XI: Results on the IMDB and AG News datasets.

Method	CIFAR-10		CIFAR-100
Method	Memory	Time	Memory	Time
	(MB)	(s)	(MB)	(s)
ResNet-110	1,564	23.81	1,564	23.93
w/ NS	1,588 ( $\uparrow$ 1.5% )	26.24 ( $\uparrow$ 10.2% )	1,588 ( $\uparrow$ 1.5% )	26.97 ( $\uparrow$ 12.7% )

TABLE XII: Computational overhead of NS in terms of GPU memory usage and per-epoch training time.

Operation	CIFAR-10	CIFAR-100
Stitching & scaling	0.14 s	0.14 s
Model inference	2.31 s	2.87 s

TABLE XIII: Per-step runtime breakdown of NS.

Panel of Figure 18 — Fig. 4: Correlation analysis on the CIFAR-LT-10/100 and Office-Home datasets. Blue points represent the individual categories, while the red line depicts the fitted linear regression.

Instance Reweighting

See all 15 papers

A Hierarchical Importance-Guided Multi-objective Evolutionary Framework for Deep Neural Network Pruning
Apr 1, 2026Zak Khan, Azam Asilian BidgoliNeural Network Pruning Neural Network Compression
Department of Computer Science and Physics, Wilfrid Laurier University, Waterloo, N2L 3C5, Ontario, Canada
Data Agent: Learning to Select Data via End-to-End Dynamic Optimization
Mar 8, 2026Suorong Yang, Fangjian Su, Hai Gan +5Data Selection Training Data Selection
National University of Singapore · Nanjing University · Nanyang Technological University
Adaptive Data Dropout: Towards Self-Regulated Learning in Deep Neural Networks
Apr 14, 2026Amar Gahir, Varshil Patel, Shreyank N GowdaTraining Data Selection Adaptive Sampling
School of Computer Science, University of Nottingham, Nottingham, United Kingdom

Apr 1, 2026J/K move · Enter open · S save

A Hierarchical Importance-Guided Multi-objective Evolutionary Framework for Deep Neural Network Pruning

Zak Khan, Azam Asilian Bidgoli

Department of Computer Science and Physics, Wilfrid Laurier University, Waterloo, N2L 3C5, Ontario, Canada

Neural Network Pruning Neural Network Compression Multi-Objective Evolutionary Optimization

Open paper AskPDF arXiv

The optimization of over-parameterized deep neural networks represents a large-scale, high-dimensional, and strongly non-convex decision problem that challenges existing optimization frameworks. Current evolutionary and gradient-based pruning methods often struggle to scale to such dimensionalities, as they rely on flat search spaces, scalarized objectives, or repeated retraining, leading to premature convergence and prohibitive computational cost. This paper introduces a hierarchical importance-guided evolutionary framework that reformulates convolutional network pruning as a tractable large-scale multi-objective optimization problem. In the first phase, a continuous evolutionary search performs coarse exploration of weight-wise pruning thresholds to shrink the search space and identify promising regions of the Pareto set. The second phase applies a fine-grained binary evolutionary optimization constrained to the surviving weights, where importance-aware sampling and adaptive variation operators refine local search in the sparse region of the Pareto set. This hierarchical design combines global exploration and localized exploitation to achieve a well-distributed Pareto set of networks balancing compactness and accuracy. Empirical results on CIFAR-10 and CIFAR-100 using ResNet-56 and ResNet-110 confirm the method's effectiveness compared to existing evolutionary approaches: pruning achieves up to 51.9% and 38.9% parameter reductions with almost no accuracy loss compared to state-of-the-art evolutionary DNN pruning methods. The proposed method contributes a scalable evolutionary approach for solving very-large-scale multi-objective optimization problems, offering a general paradigm extendable to other domains where the decision space is exponentially large, objective functions are conflicting, and efficient trade-off discovery is essential.

Abstract

Instance Reweighting

Explore similar work

A Hierarchical Importance-Guided Multi-objective Evolutionary Framework for Deep Neural Network Pruning

Data Agent: Learning to Select Data via End-to-End Dynamic Optimization

Adaptive Data Dropout: Towards Self-Regulated Learning in Deep Neural Networks