QSFM: Model Pruning Based on Quantified Similarity Between Feature Maps for AI on Edge

Wang, Zidu; Liu, Xuexin; Huang, Liechao; Chen, Yunqing; Zhang, Yufei; Lin, Zhikang; Wang, Rui

doi:10.1109/jiot.2022.3190873

Cited by 15 publications

(6 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In the scenario involving significant compression, with approximately 70% compression, CORING outperforms a recent SOTA method [34] in all aspects. FPAC [76] 93.66 0.39M(61.9) 113.08M(59.9) HRank-2 [38] 93.68 0.48M(53.8) 110.15M(61.0) EZCrop [40] 93.76 0.39M(61.9) 113.08M(59.9) DECORE-70 [1] 94.04 0.37M(65.0) 128.13M(54.7) CORING (Ours) 94.30 0.45M(57.3) 134.86M(53.5) QSFM-PSNR [73] 92.06 1.67M(25.4) 57.27M(39.4) DMC [15] 94.49 N/A 56.72M(40.0) SCOP [66] 94.24 1.43M(36.1) 56.44M(40.3) GFBS [45] 94.25 N/A 54.83M(42.0) CORING (Ours) 94.44 0.77M(65.6) 38.00M(60.0) DenseNet-40. Managing DenseNet architecture can be challenging because removing a single channel from the architecture requires removing that channel from all subsequent layers [1].…”

Section: Results and Analysismentioning

confidence: 99%

“…Comparison. CORING is compared with 44 SOTAs in the fields of structured pruning [1,3,5,6,7,11,12,15,16,18,19,23,26,29,30,31,32,34,35,36,37,38,39,40,41,43,45,46,47,48,49,56,58,63,64,66,68,73,76,78,79,80,83,85]. For a fair comparison, all available baseline models are identical.…”

Section: Experimental Settingsmentioning

confidence: 99%

“…However, methods that rely only on individual filter information neglect the correlation between filters, leading to redundancy in the resulting pruned models. Recently, similarity-based filter pruning [62,71,73,79] has become an emerging approach to eliminate redundant filters that contribute similarly to the final output. To achieve this, filters are compared to each other based on a similarity metric, and those with the highest similarity are removed during the pruning process.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Efficient tensor decomposition-based filter pruning

Pham,

Zniyed,

Nguyen

2024

Neural Networks

View full text Add to dashboard Cite

Section: Results and Analysismentioning

confidence: 99%

Section: Experimental Settingsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Efficient tensor decomposition-based filter pruning

Pham,

Zniyed,

Nguyen

2024

Neural Networks

View full text Add to dashboard Cite

“…This is much faster than the average speed of ROI generation by an IPP deployed nearshore according to our experience (Li et al, 2022). In the future, the model can be light-weighted by network pruning (Molchanov et al, 2019), quantification (Wang et al, 2022), knowledge distillation (Yim et al, 2017) techniques to further reduce its demand on computation for training and inference. These will enable the deployment of IsPlanktonFE on a cloud computing platform to facilitate next-generation in situ real-time marine plankton observations.…”

Section: Impact On Marine Plankton Observationmentioning

confidence: 93%

Deep focus-extended darkfield imaging for in situ observation of marine plankton

Chen

et al. 2023

Front. Mar. Sci.

View full text Add to dashboard Cite

Darkfield imaging can achieve in situ observation of marine plankton with unique advantages of high-resolution, high-contrast and colorful imaging for plankton species identification, size measurement and abundance estimation. However, existing underwater darkfield imagers have very shallow depth-of-field, leading to inefficient seawater sampling for plankton observation. We develop a data-driven method that can algorithmically refocus planktonic objects in their defocused darkfield images, equivalently achieving focus-extension for their acquisition imagers. We devise a set of dual-channel imaging apparatus to quickly capture paired images of live plankton with different defocus degrees in seawater samples, simulating the settings as in in situ darkfield plankton imaging. Through a series of registration and preprocessing operations on the raw image pairs, a dataset consisting of 55 000 pairs of defocused-focused plankter images have been constructed with an accurate defocus distance label for each defocused image. We use the dataset to train an end-to-end deep convolution neural network named IsPlanktonFE, and testify its focus-extension performance through extensive experiments. The experimental results show that IsPlanktonFE has extended the depth-of-field of a 0.5× darkfield imaging system to ~7 times of its original value. Moreover, the model has exhibited good content and instrument generalizability, and considerable accuracy improvement for a pre-trained ResNet-18 network to classify defocused plankton images. This focus-extension technology is expected to greatly enhance the sampling throughput and efficiency for the future in situ marine plankton observation systems, and promote the wide applications of darkfield plankton imaging instruments in marine ecology research and aquatic environment monitoring programs.

show abstract

“…In addition, computing the statistical information of the next layer and minimizing the feature reconstruction error is also a popular idea [21,26]. Different from methods that are based on only a single filter or feature, [27,23,28] remove redundant parts of a model by measuring the difference between the filters or features.…”

Section: Related Workmentioning

confidence: 99%

Network Pruning via Feature Shift Minimization

Duan¹,

Zhou²,

He³

et al. 2022

Preprint

View full text Add to dashboard Cite

Channel pruning is widely used to reduce the complexity of deep network models. Recent pruning methods usually identify which parts of the network to discard by proposing a channel importance criterion. However, recent studies have shown that these criteria do not work well in all conditions. In this paper, we propose a novel Feature Shift Minimization (FSM) method to compress CNN models, which evaluates the feature shift by converging the information of both features and filters. Specifically, we first investigate the compression efficiency with some prevalent methods in different layer-depths and then propose the feature shift concept. Then, we introduce an approximation method to estimate the magnitude of the feature shift, since it is difficult to compute it directly. Besides, we present a distribution-optimization algorithm to compensate for the accuracy loss and improve the network compression efficiency. The proposed method yields state-of-the-art performance on various benchmark networks and datasets, verified by extensive experiments. The codes can be available at https://github.com/lscgx/FSM.

show abstract

QSFM: Model Pruning Based on Quantified Similarity Between Feature Maps for AI on Edge

Cited by 15 publications

References 28 publications

Efficient tensor decomposition-based filter pruning

Efficient tensor decomposition-based filter pruning

Deep focus-extended darkfield imaging for in situ observation of marine plankton

Network Pruning via Feature Shift Minimization

Contact Info

Product

Resources

About