Maintained by Difan Deng and Marius Lindauer.
The following list considers papers related to neural architecture search. It is by no means complete. If you miss a paper on the list, please let us know.
Please note that although NAS methods steadily improve, the quality of empirical evaluations in this field are still lagging behind compared to other areas in machine learning, AI and optimization. We would therefore like to share some best practices for empirical evaluations of NAS methods, which we believe will facilitate sustained and measurable progress in the field. If you are interested in a teaser, please read our blog post or directly jump to our checklist.
Transformers have gained increasing popularity in different domains. For a comprehensive list of papers focusing on Neural Architecture Search for Transformer-Based spaces, the awesome-transformer-search repo is all you need.
0000
14.
Changwei, Ping Haoyu Ma Yongjie Song
Evolutionary Neural Architecture Search for Traffic Sign Recognition Journal Article
In: 光电子快报, 0000.
@article{nokey,
title = {Evolutionary Neural Architecture Search for Traffic Sign Recognition},
author = {Ping Haoyu Ma Yongjie Song Changwei},
url = {http://www.oelett.net/gdzkb/article/abstract/2024067},
journal = {光电子快报},
publisher = {the Editorial Department of Optoelectronics Letters},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
13.
Hosseinzadeh, Hamed
Neural Architecture Search for Adaptive Neural Network Structures: Comparative Analysis of Layer and Neuron Adjustments Miscellaneous
0000.
@misc{Hosseinzadeh-misc241,
title = {Neural Architecture Search for Adaptive Neural Network Structures: Comparative Analysis of Layer and Neuron Adjustments},
author = {Hamed Hosseinzadeh},
url = {https://www.researchsquare.com/article/rs-4909959/v1},
keywords = {},
pubstate = {published},
tppubtype = {misc}
}
12.
Deevi, Sri Aditya; Mishra, Asish Kumar; Mishra, Deepak; Kumar, L Ravi; Kumar, G V P Bharat; Bhagavan, G Murali Krishna
Efficient Self-Supervised Neural Architecture Search Booklet
EasyChair Preprint 14726, 0000.
@booklet{EasyChair:14726,
title = {Efficient Self-Supervised Neural Architecture Search},
author = {Sri Aditya Deevi and Asish Kumar Mishra and Deepak Mishra and L Ravi Kumar and G V P Bharat Kumar and G Murali Krishna Bhagavan},
url = {https://easychair.org/publications/preprint/6hNw},
howpublished = {EasyChair Preprint 14726},
month = {00},
keywords = {},
pubstate = {published},
tppubtype = {booklet}
}
11.
(Ed.)
An Efficient Neural Architecture Search Model for Medical Image Classification Collection
0000.
@collection{xie-esann24a,
title = {An Efficient Neural Architecture Search Model for Medical Image Classification},
author = {Lunchen Xie and Eugenio Lomurno and Matteo Gambella and Danilo Ardagna and Manuel Roveri and Matteo Matteucci and and Qingjiang Shi},
url = {https://www.esann.org/sites/default/files/proceedings/2024/ES2024-119.pdf},
booktitle = {European Symposium on Artificial Neural Networks 2024},
keywords = {},
pubstate = {published},
tppubtype = {collection}
}
10.
Banerjee, Jishnu
Autonomous Task-Tiling and Deep Neural Architecture Search for Intermittent Systems PhD Thesis
0000.
@phdthesis{nokey,
title = {Autonomous Task-Tiling and Deep Neural Architecture Search for Intermittent Systems},
author = {Banerjee, Jishnu},
url = {https://www.proquest.com/openview/0709fcf82e103976505463c37c325677/1?pq-origsite=gscholar&cbl=18750&diss=y},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
9.
(Ed.)
Educational Evolutionary Neural Architecture Search for Time Series Prediction Collection
0000.
@collection{nokey,
title = {Educational Evolutionary Neural Architecture Search for Time Series Prediction},
author = {Martha Isabel Escalona-Llaguno and Sergio M. Sarmiento-Rosales},
url = {https://www.scitepress.org/Papers/2024/129489/129489.pdf},
doi = {10.5220/0012948900003837},
booktitle = {Proceedings of the 16th International Joint Conference on Computational Intelligence (IJCCI 2024)},
keywords = {},
pubstate = {published},
tppubtype = {collection}
}
8.
Kim, Youngeun
Algorithmic Approaches for Empowering Spike-based Machine Intelligence PhD Thesis
0000.
@phdthesis{kim-phds25a,
title = {Algorithmic Approaches for Empowering Spike-based Machine Intelligence},
author = {Youngeun Kim
},
url = {https://www.proquest.com/docview/3164060689?pq-origsite=gscholar&fromopenview=true&sourcetype=Dissertations%20&%20Theses},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
7.
García, Cosijopii García
Multi-objective Evolutionary Algorithms for the optimization of Deep Neural Network Architectures PhD Thesis
0000.
@phdthesis{nokey,
title = {Multi-objective Evolutionary Algorithms for the optimization of Deep Neural Network Architectures},
author = {Cosijopii García García},
url = {https://inaoe.repositorioinstitucional.mx/jspui/bitstream/1009/2656/1/GARC%C3%8DAGC_DCC.pdf},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
6.
Rafiq, Shehla; Assad, Assif
RNAS-sgRNA: Recurrent Neural Architecture Search for Detection of On-Target Effects in Single Guide RNA Journal Article
In: Journal of Computational Biology, vol. 0, no. 0, pp. null, 0000, (PMID: 40501348).
@article{doi:10.1089/cmb.2025.0031,
title = {RNAS-sgRNA: Recurrent Neural Architecture Search for Detection of On-Target Effects in Single Guide RNA},
author = {Shehla Rafiq and Assif Assad},
url = {https://doi.org/10.1089/cmb.2025.0031},
doi = {10.1089/cmb.2025.0031},
journal = {Journal of Computational Biology},
volume = {0},
number = {0},
pages = {null},
abstract = {Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)/Cas9 is a leading genomic editing tool, but its effectiveness is limited by considerable heterogeneity in target efficiency among different single guide RNAs (sgRNA). This study presents RNAS-sgRNA, a hybrid model that integrates neural architecture search (NAS) with recurrent neural networks (RNN) to evaluate the on-target efficacy of CRISPR/Cas9 sgRNA. The RNAS-sgRNA model automates architectural discovery, improving sgRNA sequence categorization without considerable manual adjustment. The NAS component improves the RNN architecture, which analyzes sgRNA sequences represented as binary matrices and produces a classification score. Upon evaluation across several datasets, RNAS-sgRNA exhibits substantial performance enhancements with multiple cell lines, comparing its area under the receiver operating characteristic curve (AUROC) performance to the baseline CRISPRpred(SEQ) and DeepCRISPR models. RNAS-sgRNA demonstrated substantial improvements in AUROC performance in several cell lines compared with existing models. Notable improvements include enhancements of 8.62% for HCT116, 121.57% for HEK293T, 13.40% for HeLa, and 20.78% for HL60 cell lines, resulting in an overall improvement of 13.46%. Compared with DeepCRISPR, the model achieved additional AUROC gains in all cell lines tested, with an average improvement of 14.74%. The study also highlighted the ability of the model to deliver superior performance on smaller datasets through transfer learning, underscoring its potential applications in personalized medicine and genetic research. RNAS-sgRNA introduces a novel integration of NAS with RNN to evaluate the efficacy of CRISPR/Cas9 sgRNA. Unlike traditional methods that require significant manual adjustments, this model automates architectural discovery, optimizing the RNN structure for sgRNA sequence analysis. Furthermore, the application of transfer learning to fine-tune the pretrained model on small cell-line datasets represents a pioneering approach in the domain. The model’s demonstrated ability to significantly outperform existing algorithms, including CRISPRpred(SEQ) and DeepCRISPR, across multiple cell lines highlights its innovative contribution to genomic editing research and personalized medicine.},
note = {PMID: 40501348},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)/Cas9 is a leading genomic editing tool, but its effectiveness is limited by considerable heterogeneity in target efficiency among different single guide RNAs (sgRNA). This study presents RNAS-sgRNA, a hybrid model that integrates neural architecture search (NAS) with recurrent neural networks (RNN) to evaluate the on-target efficacy of CRISPR/Cas9 sgRNA. The RNAS-sgRNA model automates architectural discovery, improving sgRNA sequence categorization without considerable manual adjustment. The NAS component improves the RNN architecture, which analyzes sgRNA sequences represented as binary matrices and produces a classification score. Upon evaluation across several datasets, RNAS-sgRNA exhibits substantial performance enhancements with multiple cell lines, comparing its area under the receiver operating characteristic curve (AUROC) performance to the baseline CRISPRpred(SEQ) and DeepCRISPR models. RNAS-sgRNA demonstrated substantial improvements in AUROC performance in several cell lines compared with existing models. Notable improvements include enhancements of 8.62% for HCT116, 121.57% for HEK293T, 13.40% for HeLa, and 20.78% for HL60 cell lines, resulting in an overall improvement of 13.46%. Compared with DeepCRISPR, the model achieved additional AUROC gains in all cell lines tested, with an average improvement of 14.74%. The study also highlighted the ability of the model to deliver superior performance on smaller datasets through transfer learning, underscoring its potential applications in personalized medicine and genetic research. RNAS-sgRNA introduces a novel integration of NAS with RNN to evaluate the efficacy of CRISPR/Cas9 sgRNA. Unlike traditional methods that require significant manual adjustments, this model automates architectural discovery, optimizing the RNN structure for sgRNA sequence analysis. Furthermore, the application of transfer learning to fine-tune the pretrained model on small cell-line datasets represents a pioneering approach in the domain. The model’s demonstrated ability to significantly outperform existing algorithms, including CRISPRpred(SEQ) and DeepCRISPR, across multiple cell lines highlights its innovative contribution to genomic editing research and personalized medicine.
5.
More, Shraddha Subhash; Bansode, Rajesh
FCN-YOLOS: An Effective Deep-Learning Model for Real-Time Object Detection Journal Article
In: Journal of Field Robotics, vol. n/a, no. n/a, 0000.
@article{https://doi.org/10.1002/rob.70001,
title = {FCN-YOLOS: An Effective Deep-Learning Model for Real-Time Object Detection},
author = {Shraddha Subhash More and Rajesh Bansode},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.70001},
doi = {https://doi.org/10.1002/rob.70001},
journal = {Journal of Field Robotics},
volume = {n/a},
number = {n/a},
abstract = {ABSTRACT Real-time object recognition is a significant field of research with numerous applications, including object tracking, video surveillance, and autonomous driving. This identifies the smallest bounding boxes that encompass the objects of interest within the input images. Nevertheless, these approaches face challenges, like limited support for quantization and suboptimal trade in achieving accurate object detection. To address these issues, a novel approach called Faster region-based Convoluted Non-monopolize search You Only Look Once neural architecture Search (FCN-YOLOS) is introduced for object detection. This approach merges the advanced feature abstraction abilities of Faster R-CNN with the efficient object recognition strengths of YOLOv8, enhanced by NAS optimization. YOLOv8 is employed for its rapid and accurate real-time detection of abandoned items, while Faster R-CNN contributes sophisticated feature extraction by utilizing statistical, grid, and Histogram of Oriented Optical Flow (HOOF) features to improve object representation and classification. Additionally, NAS optimizes hyperparameters by balancing exploration and exploitation, which helps minimize the loss function, reduce overfitting, and enhance generalization. This results in exceptional real-time object detection performance within the FCN-YOLOS framework. The proposed technique has demonstrated a maximum image of approximately 99%, 96.3%, 94.9%, and 95.2% concerning brightness realization compared to existing methods for accuracy, recall, precision, and F1 score, respectively. These outcomes highlight its extensive applicability across diverse object detection contexts, rendering it a compelling option for both academic and industrial research. Overall, the proposed approach for object recognition techniques in feature extraction and hyperparameter adjustments further improves evaluation in terms of efficiency and object detection accuracy.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
ABSTRACT Real-time object recognition is a significant field of research with numerous applications, including object tracking, video surveillance, and autonomous driving. This identifies the smallest bounding boxes that encompass the objects of interest within the input images. Nevertheless, these approaches face challenges, like limited support for quantization and suboptimal trade in achieving accurate object detection. To address these issues, a novel approach called Faster region-based Convoluted Non-monopolize search You Only Look Once neural architecture Search (FCN-YOLOS) is introduced for object detection. This approach merges the advanced feature abstraction abilities of Faster R-CNN with the efficient object recognition strengths of YOLOv8, enhanced by NAS optimization. YOLOv8 is employed for its rapid and accurate real-time detection of abandoned items, while Faster R-CNN contributes sophisticated feature extraction by utilizing statistical, grid, and Histogram of Oriented Optical Flow (HOOF) features to improve object representation and classification. Additionally, NAS optimizes hyperparameters by balancing exploration and exploitation, which helps minimize the loss function, reduce overfitting, and enhance generalization. This results in exceptional real-time object detection performance within the FCN-YOLOS framework. The proposed technique has demonstrated a maximum image of approximately 99%, 96.3%, 94.9%, and 95.2% concerning brightness realization compared to existing methods for accuracy, recall, precision, and F1 score, respectively. These outcomes highlight its extensive applicability across diverse object detection contexts, rendering it a compelling option for both academic and industrial research. Overall, the proposed approach for object recognition techniques in feature extraction and hyperparameter adjustments further improves evaluation in terms of efficiency and object detection accuracy.
4.
Dantas, Pierre V.; Cordeiro, Lucas C.; Junior, Waldir S. S.
A Review of State-of-the-Art Techniques for Large Language Model Compression Journal Article
In: Complex & Intelligent Systems, 0000, ISSN: 2198-6053.
@article{39081e27cda444049c39fff371ed2d60,
title = {A Review of State-of-the-Art Techniques for Large Language Model Compression},
author = {Pierre V. Dantas and Lucas C. Cordeiro and Waldir S. S. Junior},
url = {https://research.manchester.ac.uk/en/publications/a-review-of-state-of-the-art-techniques-for-large-language-model-},
issn = {2198-6053},
journal = {Complex & Intelligent Systems},
publisher = {Springer Nature},
abstract = {The rapid advancement of large language models (LLMs) has driven significant progress in natural language processing (NLP) and related domains. However, their deployment remains constrained by challenges related to computation, memory, and energy efficiency – particularly in real-world applications. This work presents a comprehensive review of state-of-the-art compression techniques, including pruning, quantization, knowledge distillation, and neural architecture search (NAS), which collectively aim to reduce model size, enhance inference speed, and lower energy consumption while maintaining performance. A robust evaluation framework is introduced, incorporating traditional metrics, such as accuracy and perplexity (PPL), alongside advanced criteria including latency-accuracy trade-offs, parameter efficiency, multi-objective Pareto optimization,and fairness considerations. This study further highlights trends and challenges, such as fairness-aware compression, robustness against adversarial attacks, and hardware-specific optimizations. Additionally, NAS-driven strategies are explored as a means to design task-aware, hardware-adaptive architectures that enhance LLM compression efficiency. Hybrid and adaptive methods are also examined to dynamically optimize computational efficiency across diverse deployment scenarios. This work not only synthesizes recent advancements and identifies open problems but also proposes a structured research roadmap to guide thedevelopment of efficient, scalable, and equitable LLMs. By bridging the gap between compression research and real-world deployment, this study offers actionable insights for optimizing LLMs across a range of environments, including mobile devices and large-scale cloud infrastructures.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
The rapid advancement of large language models (LLMs) has driven significant progress in natural language processing (NLP) and related domains. However, their deployment remains constrained by challenges related to computation, memory, and energy efficiency – particularly in real-world applications. This work presents a comprehensive review of state-of-the-art compression techniques, including pruning, quantization, knowledge distillation, and neural architecture search (NAS), which collectively aim to reduce model size, enhance inference speed, and lower energy consumption while maintaining performance. A robust evaluation framework is introduced, incorporating traditional metrics, such as accuracy and perplexity (PPL), alongside advanced criteria including latency-accuracy trade-offs, parameter efficiency, multi-objective Pareto optimization,and fairness considerations. This study further highlights trends and challenges, such as fairness-aware compression, robustness against adversarial attacks, and hardware-specific optimizations. Additionally, NAS-driven strategies are explored as a means to design task-aware, hardware-adaptive architectures that enhance LLM compression efficiency. Hybrid and adaptive methods are also examined to dynamically optimize computational efficiency across diverse deployment scenarios. This work not only synthesizes recent advancements and identifies open problems but also proposes a structured research roadmap to guide thedevelopment of efficient, scalable, and equitable LLMs. By bridging the gap between compression research and real-world deployment, this study offers actionable insights for optimizing LLMs across a range of environments, including mobile devices and large-scale cloud infrastructures.
3.
Cassimon, Amber
0000.
@phdthesis{Cassimon-25a,
title = { Building efficient neural networks using scalable and transferable neural architecture search strategies },
author = { Cassimon, Amber },
url = {https://repository.uantwerpen.be/desktop/irua},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
2.
Yu, Jiaying; Luo, Haoqiu; Dai, Xinfa; Wu, Yong; Cong, Peijin; Zhou, Junlong
A Multi-Level Lightweight Framework for Deep Neural Networks on Resource-Constrained Edge Devices Journal Article
In: Journal of Circuits, Systems and Computers, vol. 0, no. ja, pp. null, 0000.
@article{doi:10.1142/S0218126626500283,
title = {A Multi-Level Lightweight Framework for Deep Neural Networks on Resource-Constrained Edge Devices},
author = {Jiaying Yu and Haoqiu Luo and Xinfa Dai and Yong Wu and Peijin Cong and Junlong Zhou},
url = {https://doi.org/10.1142/S0218126626500283},
doi = {10.1142/S0218126626500283},
journal = {Journal of Circuits, Systems and Computers},
volume = {0},
number = {ja},
pages = {null},
abstract = {Deep neural networks (DNNs) are widely used in Artificial Intelligence (AI) applications due to their powerful representational capacity. However, deploying DNNs on edge devices faces severe challenges, including limited computational resources, restricted memory capacity, and heterogeneous processor architectures, which often make conventional DNN models infeasible. In this paper, we propose a novel multi-level model lightweighting framework designed to address these challenges. The framework integrates three key techniques: Neural Architecture Search (NAS), Knowledge Distillation (KD), and layer-adaptive quantization. NAS is employed to automatically discover an efficient network architecture under hardware constraints, enabling structural compression. Subsequently, KD transfers knowledge from the original model to the compressed model, mitigating the accuracy loss resulting from compression. Finally, a layer-adaptive quantization strategy is implemented, assigning distinct bit-widths to different layers based on their data distribution characteristics, thereby enhancing storage and computational efficiency without sacrificing performance. Experimental evaluations on edge devices with multiple different processor architectures demonstrate the efficacy of the framework, achieving a model compression rate ranging from 15% to 70% with less than a 1% reduction in accuracy.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Deep neural networks (DNNs) are widely used in Artificial Intelligence (AI) applications due to their powerful representational capacity. However, deploying DNNs on edge devices faces severe challenges, including limited computational resources, restricted memory capacity, and heterogeneous processor architectures, which often make conventional DNN models infeasible. In this paper, we propose a novel multi-level model lightweighting framework designed to address these challenges. The framework integrates three key techniques: Neural Architecture Search (NAS), Knowledge Distillation (KD), and layer-adaptive quantization. NAS is employed to automatically discover an efficient network architecture under hardware constraints, enabling structural compression. Subsequently, KD transfers knowledge from the original model to the compressed model, mitigating the accuracy loss resulting from compression. Finally, a layer-adaptive quantization strategy is implemented, assigning distinct bit-widths to different layers based on their data distribution characteristics, thereby enhancing storage and computational efficiency without sacrificing performance. Experimental evaluations on edge devices with multiple different processor architectures demonstrate the efficacy of the framework, achieving a model compression rate ranging from 15% to 70% with less than a 1% reduction in accuracy.
1.
Wang, Haishuai; Gao, Yang; Zheng, Xin; Zhang, Peng; Bu, Jiajun; Yu, Philip S.
Graph neural architecture search with large language models Journal Article
In: Science China Information Sciences , 0000.
@article{nokey,
title = {Graph neural architecture search with large language models},
author = {
Haishuai Wang and Yang Gao and Xin Zheng and Peng Zhang and Jiajun Bu and Philip S. Yu
},
url = {https://link.springer.com/article/10.1007/s11432-024-4539-1},
journal = {Science China Information Sciences },
keywords = {},
pubstate = {published},
tppubtype = {article}
}