Indonesian J our nal of Electrical Engineering and Computer Science V ol. 43, No. 1, July 2026, pp. 314 ∼ 324 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v43.i1.pp314-324 ❒ 314 Lay er -wise adapti v e structur ed pruning via genetic algorithms with taylor -based pr oxy tness Anh-T ruong V o 1,2 , Hoang-Loc T ran 1,2 , Dinh-Duy Phan 1,2 , Duc-Lung V u 1,2 1 F aculty of Computer Engineering, Uni v ersity of Information T echnology , Ho Chi Minh City , V ietnam 2 V ietnam National Uni v ersity , Ho Chi Minh City , V ietnam Article Inf o Article history: Recei v ed Feb 7, 2026 Re vised Jun 6, 2026 Accepted Jun 27, 2026 K eyw ords: Layer collapse Model compression Multi-objecti v e optimization Neural architecture search NSGA-II Structured pruning T aylor e xpansion ABSTRA CT Deplo ying deep con v olutional neural netw orks (CNNs) on edge de vices requires balancing model accurac y and computational ef cienc y . While structured prun- ing limits inference costs by remo ving re dundant lters, most methods apply a rigid, global criterion, ignoring t he distinct representational roles of indi vidual layers. This yields suboptimal results, especially under aggressi v e compres- sion where o v er -pruning de grades performance . T o address this limitation, we propose an adapti v e structured pruning frame w ork based on genetic algorithms (GAs) that jointly optimizes layer -wise pruning ratios and strate gies. Each layer independently selects between min-importance and median-rank pruning, en- abling the e xploration of tail ored strate gy combinations. A training-free taylor - based proxy tness function ensures ef ci ent candidate e v aluation without re- peated ne-tuning. Aft er ne-tuning the selected architecture, e xperiments on V GG16 demonstrate that our method achie v es 92.78 ± 0.28% accurac y (o v er 50 independent runs) with a 70.0 ± 3.2% MA Cs reduction on CIF AR-10, and main- tains 71.82% accurac y on CIF AR-100. These results demonstrate competiti v e performance compared to e xisting pruning methods while achie ving substantial computational cost reduction. This is an open access article under the CC BY -SA license . Corresponding A uthor: Duc-Lung V u F aculty of Computer Engineering, Uni v ersity of Information T echnology Ho Chi Minh City , V ietnam Email: lungvd@uit.edu.vn 1. INTR ODUCTION In the era of ubiquitous articial intelligence, deep con v olutional neural netw orks (CNNs) dri v e s tate- of-the-art computer vision systems. Ho we v er , popular architectures lik e V GG16 [1] and ResNet [2] demand billions of oating-point operations (FLOPs) and millions of parameters per inference. This mismatch between modern deep model comple xity and the limited resources of edge de vice constitutes a critical deplo yment bottleneck [3]-[5]. T o address these constraints, structured pruning (lter or channel pruning) has emer ged as a prominent model compression technique. By ph ysically eliminating redundant lters, it directly reduces memory footprint and inference latenc y on general-purpose hardw are without specialized sparse computation support [6]. Despite its potential, current structured pruning f aces tw o primary limitations. First, most methods utilize rigid heuristic strate gies (e.g., L1-norm [7]) or uniform pruning ratios globally . This ignores di v erse layer roles—shallo w layers capture lo w-le v el te xtures, while deeper layers encode semantic abstractions—causing suboptimal feature preserv ation. Second, e v aluation inef ci enc y plagues automated frame w orks lik e AMC [8], J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 315 which require computationally prohibiti v e ne-tuning to e v aluate candidates. Furthermore, aggressi v e global compression can completely eliminate lters in certain layers, causing layer collapse and emphasizing the need for adapti v e pruning decisions. T o o v ercome the constraints of x ed global criteria, we address adapti v e layer -wise strate gy selec tion as a lar gely undere xplored area. W e propose an adapti v e structured pruning frame w ork based on genetic algo- rithms (GAs) [9]. Our method introduces a layer -wise h ybrid strate gy space, e n a bling each layer to indepen- dently select between min-importance pruning and median-rank pruning. The pruning process is for mulated as single-objecti v e and multi-objecti v e optimization problems, e xplicitly balancing accurac y , MA Cs, and model size. T o bypass costly ne-tuning, candidate architectures are e v aluated using a training-free taylor -based proxy tness function deri v ed from pre-computed rst-order importance scores. The main contrib utions are: − A no v el layer -wise h ybrid search space: we introduce a dual-v ariable space that jointly optimizes pruning ratios and strat e gi es (min-importance vs. median-rank) per l ayer , unlik e con v entional methods conned to optimizing ratios under a x ed criterion. − T opology-preserving constrained optimization: we frame pruning as a constrained multi-objecti v e problem, enforcing structural constraints ac ross the decoding and tness e v aluation phases to maintain topological connecti vity and pre v ent layer collapse under aggressi v e compression. − Adapti v e strate gy disco v ery: our frame w ork autonomously identies h ybrid strate gies balancing ef cienc y and feature preserv ation. Ov er 50 independent runs, the multi-objecti v e v ariant achie v ed 92.78 ± 0.28% accurac y while reducing MA Cs by 70.0 ± 3.2%. T o our kno wledge, this w ork is among the rst pruning frame w orks that jointly optimize layer -wise pruning ratios and pruning strate gies using e v olutionary multi-objecti v e optimization. 2. RELA TED W ORK 2.1. Structur ed pruning f or efcient deep netw orks Among v arious model compression techniques, structured pruning has g ained particular attention for edge deplo yment [6], [10] because it ph ysically remo v es redundant lters or channels, achie ving direct re- ductions in memory footprint and inference latenc y on general-purpose hardw are without requiring sparse computation support. A central challenge in structured pruning is identifying redundant or unimportant lters. Early ap- proaches relied on heuristic metrics such as weight magnitude, with L1-norm pruning [7] assuming that lters with small weights contrib ute less to model performance. Subsequent w orks proposed data-dri v en criteria to better capture fe ature redundanc y . ThiNet [11] prunes lters bas ed on statistics from the ne xt layer , while net- w ork slimming [12] le v erages channel-wise scaling f actors from batch normalization. FPGM [13] introduces a geometric median-based criterion for lter selection. F or e xample, HRank [14] measures the rank of feature maps to estimate l ter importance, while Molchano v et al. [15], [16] introduced a theoretically grounded cri- terion based on the rs t-order taylor e xpansion of the loss function, approximated by the product of acti v ation and gradient magnitude. Despite theoretical adv ances such as the Lottery T ick et Hypothesis [17] and pruning at initiali zation [18], and the observ ation by Liu et al. [19] that pruned architectures can be retrained from scratch with compet- iti v e performance, most e xisting pruning methods still apply importance criteria in a greedy or globally uniform manner , ignoring that dif ferent layers require distinct preserv ation strate gies. 2.2. A utomated pruning via sear ch T o a v oid manual tuning of per -layer pruning ratios, recent studi es ha v e e xplored automated pruning frame w orks based on search or optimization. AMC [8] formulates s tructured pruning as a reinforce- ment learning problem, where an agent learns layer -wise compression ratios. While ef fecti v e, AMC relies on ne-tuning-based re w ard e v aluation, resulting in substantial computational cost. ABCPruner [20] further models channel pruning as an optimization-based structure search problem, enabling e xible e xploration of the pruning space. Similarly , Feng et al. [21] propose a task-dri v en sparsity approach that automatically learns layer -wise pruning rules via a dif ferentiable search mechanism, reducing the reliance on manually designed criteria. Layer -wise adaptive structur ed pruning via g enetic algorithms with taylor ... (Anh-T ruong V o) Evaluation Warning : The document was created with Spire.PDF for Python.
316 ❒ ISSN: 2502-4752 T o reduce e v aluation cost, proxy-based methods ha v e been proposed. EagleEye [22] estimates sub- netw ork accurac y without full training by le v eraging adapti v e batch normalization statistics, signicantly accel- erating the search process. MetaPruning [23] uses meta-learning to generate pruned netw ork weights directly , while once-for -all (OF A) [24] trains a single super -netw ork that can be specialized for di v erse deplo yment scenarios. Ef cient netw ork design has also been adv anced by architectures such as MobileNetV2 [25] and Ef cientNet [26], which incorporate ef cient b uilding blocks. Despite these adv ances, most automated prun- ing methods primarily focus on optimizing ho w much to prune per layer , while implicitly assuming a x ed pruning criterion (e.g., L1-norm) throughout the netw ork. Moreo v er , prior e v olutionary approaches typically formulate pruning as a single-objecti v e problem (e.g., maximizing accurac y under a FLOPs constraint). While recent w orks ha v e be gun e xploring multi- objecti v e e v olutionary pruning [27], [28], these methods primarily search for pruning ratios under a x ed, globally uniform criterion. In contrast, our w ork emplo ys multi-obj ecti v e optimization (MOO) to jointly e xplore both pruning ratios and per -layer pruning strate gies, e xplicitly trading of f accurac y , ef cienc y , and model size. 2.3. Gap analysis and moti v ation Despite recent adv ances in automated pruning, e xisting methods still f ace limitations in both the design of the pruning search space and the ef cienc y of candidate e v aluation during the search process. − Rigidity of global pruning criteria: while automated frame w orks lik e AMC [8] and ABCPruner [20] can search for distinct layer -wise ratios, the y typically rely on a globally uniform pruning criterion across all layers. Such a uniform pruning criterion may not fully account for the heterogeneous representational roles of shallo w and deep layers. Consequently , the search process often focuses primari ly on optimizing ho w much to prune, while the choice of ho w to prune is usually predened. − High search latenc y and e v aluation challenges: e v olutionary and re inforcement learning-based pruning methods often incur high computational costs due to repeated ne-tuning or v alidation c ycles. Proxy-based strate gies, such as EagleEye [22], mitig ate this b urden by estimating sub-netw ork performance without full retraining. Ho we v er , these approaches typically require a calibration phase to recompute batch normaliza- tion statist ics through se v eral forw ard passes on a calibration set for each candidate sub-netw ork. While relati v ely lightweight, this additional s tep can introduce non-ne gligible cumulati v e latenc y when e v aluating a lar ge population of candidates in e xtensi v e search spaces. Our moti v ation. These li mitations moti v ate the de v elopment of a frame w ork that jointly optimizes ho w much to prune and ho w to prune at each layer . By inte grating a h ybrid strate gy space with a training- free T aylor -based proxy , we aim to pro vide an ef cient, adapti v e, and topology-preserving solution for model compression. 3. METHOD T o address CNN resource optimization, this study rst in v estig ates global pruning strate gies to identify structural limitations, then proposes a GA frame w ork [9] that jointly searches layer -wise pruning ratios and pruning strate gies. 3.1. Pr eliminary study on global strategies W e empirically e v aluate global structured pruning to analyze the impact of lter selection c riteria. The importance score I ( f i ) for each lter f i in V GG16 relies on the rst-order taylor e xpansion [15], [16]. W e e v aluated three selection strate gies: min-importance prunes lo west-score lters to preserv e sensi- ti v e features; max-importance prunes highest-score lters (h ypothesizing the y cause o v ertting); and median- rank preserv es lters representing the central tendenc y ⌊ N − k 2 ⌋ , ⌊ N + k 2 ⌋ . Our proposed h ybrid strate gy balances immediate accurac y with global netw ork inte grity . While min- importance maintains accurac y in high-v ariance layers, its rigid application in deep, attenuated layers causes layer collapse . Consequently , median-rank serv es as a distrib utional re gularizer , maintaining feature di v ersity and topological connecti vity . This dual space lets the GA adapti v ely determine optimal layer -wise trade-of fs. Quantitati v e analysis of layer -wise score imbalance: T o understand the f ailure of global thresholds in deep layers, we analyzed the statistical distrib ution of taylor importance scores across all 13 con v olutional layers of V GG16 (T able 1). Indonesian J Elec Eng & Comp Sci, V ol. 43, No. 1, July 2026: 314–324 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 317 T able 1. Distrib ution of taylor scores across all V GG16 con v olutional layers (calculated via a representati v e pre-computation pass on a CIF AR-10 data batch) Layer M ax score Min score Mean score Std. De v Attenuation (vs. L00 Peak) Layer 00 (Con v1 1) 0.5680 0.0001 0.0768 0.0986 1.000 (100.0%) Layer 01 (Con v1 2) 0.4023 0.0005 0.0956 0.0805 0.708 (70.8%) Layer 02 (Con v2 1) 0.3023 0.0003 0.0692 0.0550 0.532 (53.2%) Layer 03 (Con v2 2) 0.3049 0.0006 0.0680 0.0564 0.537 (53.7%) Layer 04 (Con v3 1) 0.3067 0.0001 0.0474 0.0408 0.540 (54.0%) Layer 05 (Con v3 2) 0.2568 0.0003 0.0487 0.0392 0.452 (45.2%) Layer 06 (Con v3 3) 0.2325 0.0001 0.0487 0.0392 0.409 (40.9%) Layer 07 (Con v4 1) 0.1899 0.0000 0.0334 0.0290 0.334 (33.4%) Layer 08 (Con v4 2) 0.1840 0.0001 0.0334 0.0289 0.324 (32.4%) Layer 09 (Con v4 3) 0.1737 0.0000 0.0346 0.0275 0.306 (30.6%) Layer 10 (Con v5 1) 0.5022 0.0000 0.0291 0.0332 0.884 (88.4%) Layer 11 (Con v5 2) 0.1570 0.0001 0.0354 0.0265 0.276 (27.6%) Layer 12 (Con v5 3) 0.0703 0.0256 0.0437 0.0066 0.124 (12.4%) Empirical data re v eals stark v ariance in feature importance distrib ution. High-v ariance layers: earl y layers (e.g., Layer 00) and specic intermediate layers e xhibit high max-to-mean v ariance, containing a fe w dominating lters with e xceptionally high scores ( ∼ 0.50–0.57). Signal attenuation: the deepest layer (layer 12) suf fers from e xtreme score attenuation. Its maximum taylor score ( 0 . 0703 ) is nearly 8 × lo wer than that of Layer 00, representing a relati v e attenuation of just 12 . 4 %. Its v ariance is e xtremely tight (std = 0 . 0066 ), meaning the scores are uniformly distrib uted near zero. This se v ere disparity creates a conict for global criteria. A modest global re tention threshold of 0 . 10 w ould catastrophically eliminate e v ery lter in layer 12 (max score = 0 . 0703 ), despite its critical high-le v el semantic features. This phenomenon, termed layer collapse, e xposes the limits of rigid global heuristics and moti v ates our adapti v e, multi-objecti v e search. T o address this limitation, Figure 1 illustrates the o v erall w orko w of the proposed adapti v e GA-based pruning frame w ork. Figure 2 sho ws the NSGA-II P areto front, illustrating the trade-of f between compression and accurac y used to select the optimal pruning solution. Figure 1. Ov ervie w of the proposed multi-objecti v e e v olutionary pruning frame w ork. The pipeline consists of three stages: (1) Pre-computation: lter importance scores are estimated once via T aylor e xpansion and cached as a lookup table for training-free proxy e v aluation. (2) NSGA-II Search: candidate pruning policies—each encoding layer -wise pruning ratios and strate gies (min-importance or median-rank)—are e v olv ed and e v aluated ag ainst three objecti v es: proxy tness, MA Cs, and con v olutional parameters. (3) Decision and deplo yment: a knee-point solution is selected from the resulting P areto front and ne-tuned with kno wledge distillation to produce the compact model. Layer -wise adaptive structur ed pruning via g enetic algorithms with taylor ... (Anh-T ruong V o) Evaluation Warning : The document was created with Spire.PDF for Python.
318 ❒ ISSN: 2502-4752 Figure 2. NSGA-II P areto front on CIF AR-10 (representati v e run). Each point is a non-dominated pruning conguration. The front re v eals a clear compression–accurac y trade-of f: solutions to w ard the lo wer -left achie v e aggressi v e compression at reduced proxy tness, while those to w ard the upper -right preserv e higher tness at greater parameter cost. The knee point identies the architecture of fering the best balance between ef cienc y and accurac y preserv ation, selected for subsequent ne-tuning. 3.2. Sear ch space and lay er -wise adapti v e pruning policy Unlik e most e xisting automated pruning methods that primarily search for pruning ratios, our frame- w ork introduces a dual search space that jointly optimizes pruning ratios (ho w much t o prune) and pruning strate gies (ho w to prune). Gi v en a CNN with L con v olutional layers, each chromosome z represents, z = [( r 1 , s 1 ) , ( r 2 , s 2 ) , . . . , ( r L , s L )] , (1) where r l ∈ (0 , 1] denotes the retention ratio of layer l , and s l denotes the selected pruning strate gy (min- importance or median-rank). W e e xclude max-importance since remo ving highly important lters typically leads to se v ere accurac y de gradation and contradicts standard pruning objecti v es. 3.3. T aylor -based pr oxy tness Ev aluating candidate pruning congurations via ne-tuning is computationally e xpensi v e during e v o- lutionary search. T o enable ef cient e v aluation, we emplo y a taylor -based proxy tness deri v ed from the rst-order taylor importance criterion [15], [16]. Gi v en a pruning conguration z retaining lters S l ( z ) in each layer , the global taylor preserv ation ratio is dened as, P T ( z ) = P L l =1 P i ∈ S l ( z ) T l ,i P L l =1 P C l i =1 T l ,i (2) where C l is the number of lters in layer l . This ratio measures the proportion of taylor importance preserv ed after pruning and serv es as an ef cient proxy for ranking candidate congurations during the e v olutionary search. F or the multi-objecti v e formulation, we use a layer -wise normalized v ariant to mitig ate dominance from high-importance layers: P nor m T ( z ) = 1 L L X l =1 P i ∈ S l ( z ) T l ,i P C l i =1 T l ,i (3) Indonesian J Elec Eng & Comp Sci, V ol. 43, No. 1, July 2026: 314–324 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 319 3.4. Single-objecti v e tness f ormulation T o establish a baseline, we designed a single-objecti v e genetic algorithm (GA-SO) aggre g ating objecti v es into a scalar tness function: F ( z ) = P T ( z ) − λ 1 · ϕ ( z ) − λ 2 · ψ ( z ) − λ 3 · r par am ( z ) (4) where P T ( z ) denotes the taylor -based proxy tness dened in section 3.3. The MA Cs penalty is dened as: ϕ ( z ) = max(0 , r mac − τ ) + β · max(0 , r mac − τ ) 2 (5) The structural vulnerability penalty is dened as: ψ ( z ) = X l ∈L v uln max(0 , ρ − r l ) 2 (6) In addition to this soft penalty , the full structural safe guard mechanism enforces a hard oor of 10% retention per layer (with lter counts rounded to multiples of 8 for hardw are alignment) during chromosome decoding. An y solution violating this hard boundary is rejected ( F = −∞ ). 3.5. Multi-objecti v e f ormulation and taylor -based pr oxy Rather than coalescing constraints into scalar tnesses, we formulate pruning as a multi-obj ecti v e optimization problem (MOP): min z F ( z ) = [ f acc ( z ) , f mac ( z ) , f param ( z )] (7) where f acc ( z ) = −P T ( z ) and P T ( z ) is the T aylor -based proxy tness dened in Section 3.3.. T o solv e (7), we emplo y the non-dominated sorting genetic algorithm II (NSGA-II) [29]. Indi viduals are rank ed via non-dominated sorting based on P areto dominance, where a solution z 1 dominates z 2 ( z 1 ≺ z 2 ) if it is no w orse in all objecti v es and strictly better in at least one. Di v ersity within each front is maintained using cro wding distance. During e v olution, we apply single-point crosso v er to e xchange genes between parent solutions. Mutation reinitializes ratio genes uniformly within [ min ratio , 1 . 0] and randomly reassigns strate gy genes. The o v erall optimization procedure is summarized in Algorithm 1. Algorithm 1 NSGA-II based h ybrid structured pruning 1: Input: Pre-computed taylor scores, population size N , generations G . 2: Initialize population P 0 with random pruning ratios and strate gies. 3: Ev aluate objecti v es F ( z ) for all z ∈ P 0 . 4: f or t = 0 to G − 1 do 5: Generate of fspring Q t via selection, crosso v er , and mutation. 6: Ev aluate objecti v es for all z ∈ Q t . 7: Mer ge populations R t = P t ∪ Q t . 8: Perform non-dominated sorting on R t to obtain fronts F 1 , F 2 , . . . . 9: Construct P t +1 using fronts and cro wding distance. 10: end f or 11: Output: P areto front F 1 . Knee point selection. The NSGA-II algorithm produces a P areto front F 1 containing multiple non- dominated solutions. T o obtain a single deplo yable model, we select the knee point z ∗ , dened as the solution minimizing the Euclidean distance to the ideal point in the normalized object i v e space. This strate gy pro vides a balanced trade-of f between accurac y preserv ation and resource reduction. z ∗ = arg min z ∈F 1 q ˆ f 2 acc ( z ) + ˆ f 2 mac ( z ) + ˆ f 2 par am ( z ) (8) Layer -wise adaptive structur ed pruning via g enetic algorithms with taylor ... (Anh-T ruong V o) Evaluation Warning : The document was created with Spire.PDF for Python.
320 ❒ ISSN: 2502-4752 4. RESUL TS AND DISCUSSION 4.1. Experimental setup W e e v aluate our frame w ork on the V GG16 architecture [1] us ing tw o benchmark datasets: CIF AR-10 and CIF AR-100 [30]. Baseline accurac y , parameters, and MA Cs for these datasets are detailed in subsequent result tables. The e v olutionary search operates with a population size of 50 o v er 30 generations on a single NVIDIA R TX 4090 GPU. The entire search process completes in under 15 seconds because taylor importance scores are pre-computed e xactly once. The 1,500 genetic e v aluations are performed using these cached ranking tables via v ectorized summation, eliminating iterati v e ne-tuning during search. Final architectures are ne-tuned for 150 epochs usi ng SGD with kno wledge distillation (KD) [31] ( T = 4 . 0 , α = 0 . 9 ). Full implementation details are publicly a v ailable (see data a v ailability). 4.2. Stability analysis: global vs. adapti v e pruning W e rst in v estig ate the rob ustness of dif ferent pruning strate gies under an e xtreme compression re gime (tar geting < 0 . 1 G MA Cs). T able 2 compares our adapti v e frame w ork with standard global pruning heuristics. As sho wn in T able 2, rigid global ranking strate gies such as median and max f ail under e xtreme compression, leading to layer collapse. In contrast, our adapti v e GA frame w orks consistently identify stable architectures that remain operable at similar compression le v els. T able 2. Stability comparison at e xtreme compression (CIF AR-10) Strate gy MA Cs (G) P arams (M) Acc (%) Status Baseline 0.33 33.65 93.62 - Global Min 0.07 18.15 92.53 Stable Global Median - - CRASH Layer collapse Global Max - - CRASH Layer collapse Adapti v e GA-SO 0.087 23.21 92.33 Stable Adapti v e GA-MO 0.099 ± 0.010 21.77 ± 0.81 92.78 ± 0.28 Stable 4.3. Comparison with structur ed pruning methods T able 3 compares our method with representati v e structured pruning methods that report results on V GG16/CIF AR-10. As s ho wn in T able 3, L1-Norm and HRank achie v e slightly higher absolute accurac y b ut operate at substantially lo wer compression le v els (34–54% MA Cs reduction). ABCPruner [20] reaches a comparable MA Cs reduction (73.68%) while retaining higher accurac y (93.08%), as it performs optimization- based ratio search. Ho we v er , ABCPruner searches only for per -layer pruning ratios under a x ed criterion. T able 3. Comparison with representati v e structured pruning methods on V GG16 (CIF AR-10) Method Acc (%) P arams (%) MA Cs (%) Baseline 93.62 0 0 L1-Norm [7] 93.40 64.0 34.2 HRank [14] 93.43 82.9 53.5 ABCPruner [20] 93.08 88.68 73.68 Ours (GA-SO) 92.33 31.0 73.6 Ours (GA-MO) 92.78 ± 0.28 35.3 ± 2.4 70.0 ± 3.2 By jointly optimizing pruning ratios and strate gies, our GA-MO v ariant achie v es 70.0 ± 3.2% MA Cs reduction with 92.78 ± 0.28% accurac y across 50 runs; the absolute comparisons should be interpreted cau- tiously due to dif fering training protocols. AMC [8], EagleEye [22], and MetaPruning [23] search for per -layer ratios b ut do not report V GG16/CIF AR-10 results and require substantially more computation (e.g., EagleEye ∼ 25 GPU hours vs. our < 15 seconds), while our method additionally e xplores how to prune at each layer . Under aggressi v e compression on CIF AR-100 (T able 4), Adapti v e GA-MO achie v es 71.82% accu- rac y with only a 0.96% drop, outperforming all global baselines Since L1-Norm, HRank, and ABCPruner do not report results on the V GG16/CIF AR-100 dataset in their original publications, direct comparison with these methods is not included in T able 4. Indonesian J Elec Eng & Comp Sci, V ol. 43, No. 1, July 2026: 314–324 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 321 T able 4. Performance comparison on CIF AR-100 (V GG16) Method MA Cs (G) Reduct. Acc (%) Drop (%) P arams (M) Drop (%) Baseline 0.33 0.0% 72.78 - 34.02 - Global Min 0.12 63.6% 69.91 -2.87 22.30 34.5% Global Median 0.14 57.6% 71.02 -1.76 21.45 37.0% Global Max 0.10 69.7% 68.57 -4.21 22.11 35.0% Ours (GA-MO) 0.116 64.8% 71.82 -0.96 22.52 33.8% Ours (GA-SO) 0.086 73.9% 70.68 -2.10 23.26 31.6% 4.4. Discussion Our analysis in section 3 sho ws that layer -wise score imbalance causes global median and max strat e- gies to suf fer from layer collapse. Our frame w ork a v oids this through layer -wise decision making, learning local ratios and strate gies simultaneously instead of relying on a single global threshold. Our e xperiments re v eal distinct beha vioral dif ferences between the GA v aria nts. The single-obje cti v e formulation (GA-SO), dri v en by a weighted-sum tness, con v er ges to a 100% min-importance strate gy across all layers. This greedy beha vior ma x i mizes the proxy tness b ut limi ts e xploration. In contrast, the multi- objecti v e formulation (GA-MO) maintains strate gy di v ersity throughout the search, as illustrated in the P areto front (Figure 2). Rather than forcing a single threshold, this spread enables selection of architectures tailored to dif ferent hardw are constraints. What is ne w and wh y it matters. By introducing a dual-v ariable search space that simult aneously optimizes ho w much and ho w to prune, our method e xtends prior automated pruning approaches that search only for pruning ratios under x ed criteria. This broader search space allo ws the e v olutionary process to disco v er pruning congurations that remain stable and a v oid layer collapse e v en under aggressi v e compression re gimes. Role of the structural safe guard. An ablation e xperiment conrms that disabling the safe guard (s oft penalty + hard constraints) does not de grade accurac y , acting as a cost-free safety mechanism. Operationally , the 8-lter alignment ensures hardw are-friendly channel counts, impro ving MA Cs estimation consistenc y , while the minimum retention oor pre v ents search re gions prone to collapse. 4.5. Pr oxy delity and statistical r ob ustness T o v alidate proxy reliability , we correlate proxy tness with post -ne-tuning accurac y across 32 non- dominated sub-netw orks from a single NSGA-II P areto front (each pruned and ne-tuned for 150 epochs with identical KD settings). The Spearman correlation is ρ = 0 . 8035 ( p = 3 . 08 × 10 − 8 ) and pears on r = 0 . 8272 ( p = 5 . 32 × 10 − 9 ), conrming that the proxy reliably ranks candidates across dif ferent compression re gimes, consistent with EagleEye [22]. Ov er 50 independent runs with dif ferent random seeds, the full GA-MO pipeline achie v es mean = 92.78%, std = 0.28%, mi n = 92.01%, max = 93.33%, indicating stable con v er gence to high-quality solutions independent of random initialization. 4.6. Limitations This w ork introduces limitations w arranting future in v estig ation: (i) Architecture scope: a prelim inary e v aluation on ResNet-56 (CIF AR-10) yields 90.07 ± 0.88% accurac y at 83.7 ± 1.7% MA Cs reduction o v er 88 runs (baseline: 93.44%), suggesting transferability to residual architectures. (ii) Dataset scale: v alidation is limited to CIF AR. Lar ger datasets (e.g., ImageNet) may e xhibit distinct pruning dynamics. (iii) Proxy approx- imation: the taylor proxy relies on a rst-order approximation and may miss second-order lter interactions. (i v) Strate gy space: e xpanding be yond the tw o current strate gies (e.g., geometric-median) may yield impro v ements. 5. CONCLUSION In this paper , we present an adapti v e structured pruning frame w ork using a multi-objecti v e e v ol ution- ary algorithm (NSGA-II) with a layer -wise h ybrid strate gy . By jointly optimizing pruning ratios and pruning strate gies at each layer , the proposed frame w ork mo v es be yond rigid global heuristics that apply a single cri- terion uniformly . The multi- ob j ecti v e formulation enables e xplicit e xploration of trade-of fs between accurac y , computational cost, and model size. Layer -wise adaptive structur ed pruning via g enetic algorithms with taylor ... (Anh-T ruong V o) Evaluation Warning : The document was created with Spire.PDF for Python.
322 ❒ ISSN: 2502-4752 Experimental results on V GG16 sho w that our method achie v es 92.78 ± 0.28% accurac y with 70.0 ± 3.2% MA Cs reduction on CIF AR-10 while maintaining stable architectures under aggressi v e compression. On CIF AR-100, the frame w ork retains 71.82% accurac y with a 64.8% MA Cs reduction, indicating good general- ization to more challenging classication tasks. Proxy delity analysis conrms that the T aylor -based proxy reliably ranks candidate archit ectures (strong rank correlation), and statistical v alidation o v er 50 independent runs demonstrates consistent repro- ducibility (std = 0 . 28% ). Ultimately , these empirical adv antages suggest that the proposed frame w ork holds signicant potential to f acilitate rob ust compression for mobile AI and edge de vices, reducing theoretical in- ference costs while preserving accurac y . While the frame w ork demonstrates strong empirical adv antages on V GG16, a preliminary e v aluation on ResNet-56 suggests trans ferability to residual architectures. Full v alidation on lar ger arc h i tectures and datasets remains as future w ork. Future w ork will focus on: (i) full statistical v alidation on residual architectures (ResNet-50/ImageNet) and T ransformer -based models; (ii) e xpanding the strate gy search space with geometric or acti v ation-based cri- teria; and (ii i) deplo ying compressed models on real-w orld edge de vices to measure ph ysical inference latenc y and ener gy ef cienc y . A CKNO WLEDGMENTS The authors w ould lik e to thank the MMLab, Uni v ersity of Information T echnology , V ietnam National Uni v ersity Ho Chi Minh City , for pro viding the computational resources used in this research. FUNDING INFORMA TION This research w as funded by Uni v ersity of Information T echnology , V ietnam National Uni v ersity Ho Chi Minh City under grant number D1-2026-06. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Anh-T ruong V o ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Hoang-Loc T ran ✓ ✓ ✓ ✓ ✓ ✓ Dinh-Duy Phan ✓ ✓ ✓ ✓ ✓ ✓ Duc-Lung V u ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administrati on V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT The authors declare that the y ha v e no kno wn competing nancial interests or personal relat ionships that could ha v e appeared to inuence the w ork reported in this paper . D A T A A V AILABILITY The e xperiments in this study were conducted using publicly a v ailable datasets (CIF AR-10 and CIF AR- 100). The source code implementing the proposed frame w ork is publicly a v ailable at: https://github. com/truongva1111/adaptive- ga- pruning . Indonesian J Elec Eng & Comp Sci, V ol. 43, No. 1, July 2026: 314–324 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 323 REFERENCES [1] K. Simon yan and A. Zisserman, “V ery deep con v olutional netw orks for lar ge-scale image recogniti on, ” 3rd International Conference on Learning Representations , ICLR 2015 - Conference T rack Proceedings , Apr . 2015, [Online]. A v ailable: http://arxi v .or g/abs/1409.1556 [2] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition, ” in 2016 IEEE Conference on Computer V ision and P attern Recognition (CVPR), IEEE, Jun. 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90. [3] F . N. Peccia, S. P a vlitska, T . Fleck, and O. Bringmann, “Ef cient Edge AI: deplo ying con v olutional neural netw orks on FPGA with the gemmini accelerator , ” in 2024 27th Euromicro Conference on Digital System Design (DSD) , IEEE, Aug. 2024, pp. 418–426. doi: 10.1109/DSD64264.2024.00062. [4] V . Kamath and A. Renuka, “Deep learning based object detection for resource constrained de vi ces: systematic re vie w , future trends and challenges ahead, ” Neurocomputing , v ol. 531, pp. 34–60, Apr . 2023, doi: 10.1016/j.neucom.2023.02.006. [5] M. Y . Shabir , G. T orta, and F . Damiani, “T in yML model compression: a comparati v e study of pruning and quantization on selected standard and custom neural netw orks, ” T elecommunication Systems , v ol. 88, no. 4, p. 132, Dec. 2025, doi: 10.1007/s11235-025- 01363-2. [6] Y . He and L. Xiao, “Structured pruning for deep con v olutional neural netw orks: a surv e y , ” IEEE T ransactions on P a ttern Analysis and Machine Intelligence , v ol. 46, no. 5, pp. 2900–2919, 2024, doi: 10.1109/TP AMI.2023.3334614. [7] H. Li, K. Asim, D. Igor , S. Hanan, and G. H. Peter , “Pruning lters for ef cient con vnets, ” 5th International Conference on Learning Representations, ICLR 2017 - Conference T rack Proceedings , v ol. 1608.08710, 2016. [8] Y . He, J. Lin, Z. Liu, H. W ang, L.-J. Li, and S. Han, “ AMC: AutoML for model compression and acceleration on mobile de vices, ” in Lecture Notes in Computer Science (including subseries Lecture Notes in Articial Intelligence and Lecture Notes in Bioinformatics) , v ol. 11211 LNCS, 2018, pp. 815–832. doi: 10.1007/978-3-030-01234-2 48. [9] J. H. John and J. Henry , “ Adaptation in nat ural and articial systems: an introductory analysis with applications to biology , control, and articial intelligence, ” Uni v ersity of Michig an Press, 1992. [10] H. Cheng, M. Zhang, and J. Q. Shi, “ A surv e y on deep neural netw ork pruning: taxonomy , comparison, analysis, and recom- mendations, ” IEEE T ransactions on P attern Analysis and Machine Intelligence , v ol. 46, no. 12, pp. 10558–10578, Dec. 2024, doi: 10.1109/TP AMI.2024.3447085. [11] J.-H. Luo, J. W u, and W . Lin, “ThiNet: a lter le v el pruning method for deep neural netw ork compression, ” in 2017 IEEE Interna- tional Conference on Computer V ision (ICCV) , IEEE, Oct. 2017, pp. 5068–5076. doi: 10.1109/ICCV .2017.541. [12] Z. Liu, J. Li, Z. Shen, G. Huang, S. Y an, and C. Zhang, “Learning ef cient con v olutional netw orks through netw ork slimming, ” in 2017 IEEE International Conference on Computer V ision (ICCV) , IEEE, Oct. 2017, pp. 2755–2763. doi: 10.1109/ICCV .2017.298. [13] Y . He, P . Liu, Z. W ang, Z. Hu, and Y . Y ang, “Filter pruning via geometric median for deep con v olutional neural netw orks accelera- tion, ” in 2019 IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR) , IEEE, Jun. 2019, pp. 4335–4344. doi: 10.1109/CVPR.2019.00447. [14] M. Lin et al. , “HRank: lter pruning using high-rank feature map, ” in 2020 IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR) , IEEE, Jun. 2020, pp. 1526–1535. doi: 10.1109/CVPR42600.2020.00160. [15] P . Molchano v , S. T yree, T . Karras, T . Aila, and J . Kautz, “Pruning con v olutional neural netw orks for resource ef cient inference, ” in 5th International Conference on Learning Representations, ICLR 2017 - Conference T rack Proceedings , 2017. [16] P . Molchano v , A. Mallya, S. T yree, I. Frosio, and J. Kautz, “Importance estimation for neural netw ork pruning, ” in 2019 IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR) , IEEE, Jun. 2019, pp. 11256–11264. doi: 10.1109/CVPR.2019.01152. [17] J. Frankle and M. Carbin, “The lottery tick et h ypothesis: nding sparse, trainable neural netw orks, ” 7th International Conference on Learning Representations, ICLR 2019 , 2019. [18] C. W ang, G. Zhang, and R. Grosse, “Picking winning tick ets before training by preserving gradient o w , ” in 8th International Conference on Learning Representations, ICLR 2020 , 2020. [19] Z. Liu, M. Sun, T . Zhou, G. Huang, and T . Darrell, “Rethinking the v alue of net w ork pruning, ” in 7th International Conferenc e on Learning Representations, ICLR 2019 , 2019. [20] M. Lin, R. Ji, Y . Zhang, B. Zhang, Y . W u, and Y . T ian, “Channel pruning via automatic structure search, ” in Proceedings of the T wenty-Ninth International Joint Conference on Articial Intelligence , C alifornia: International Joint Conferences on Articial Intelligence Or g anization, Jul. 2020, pp. 673–679. doi: 10.24963/ijcai.2020/94. [21] K.-Y . Feng, X. Fei, M. Gong, A. K. Qin, H. Li, and Y . W u, “ An automatically layer -wise searching strate gy for channel pruning based on task-dri v en sparsity optimization, ” IEEE T ransactions on Circuits and Systems for V ideo T echnology , v ol. 32, no. 9, pp. 5790–5802, Sep. 2022, doi: 10.1109/TCSVT .2022.3156588. [22] B. Li, B. W u, J. Su, and G. W ang, “EagleEye: f ast sub-net e v aluation for ef cient neural netw ork pruning, ” in Lecture Notes in Computer Science (including s ubseries Lecture Notes in Articial Intelligence and Lecture Notes in Bioinformatics) , v ol. 12347 LNCS, 2020, pp. 639–654. doi: 10.1007/978-3-030-58536-5 38. [23] Z. Liu et al. , “MetaPruning: meta learning for automatic neural netw ork channel pruning, ” Proceedings of the IEEE International Conference on Computer V ision , pp. 3295–3304, 2019, doi: 10.1109/ICCV .2019.00339. [24] H. Ca i, C. Gan, T . W ang, Z. Zhang, and S. Han, “Once-for -all: train one netw ork and specialize it for ef cient deplo yment, ” in 8th International Conference on Learning Representations, ICLR 2020 , 2020. [25] M. Sandler , A. Ho w ard, M. Zhu, A. Zhmogino v , and L.-C. Chen, “MobileNetV2: in v erted residuals and linear bottle- necks, ” in 2018 IEEE/CVF Conference on Computer V ision and P attern Recognition , IEEE, Jun. 2018, pp. 4510–4520. doi: 10.1109/CVPR.2018.00474. [26] M. T an and Q. V . Le, “Ef cientNet: rethinking model scaling for con v olutional neural netw orks, ” in 36th International Conference on Machine Learning, ICML 2019 , 2019, pp. 6105–6114. [27] W . Hong, G. Li, S. Liu, P . Y ang, and K. T ang, “Multi-objecti v e e v olutionary optimization for hardw are-a w are neural netw ork prun- ing, ” Fundamental Research , v ol. 4, no. 4, pp. 941–950, Jul. 2024, doi: 10.1016/j.fmre.2022.07.013. Layer -wise adaptive structur ed pruning via g enetic algorithms with taylor ... (Anh-T ruong V o) Evaluation Warning : The document was created with Spire.PDF for Python.