IoTNet: An Efficient and Accurate Convolutional Neural Network for IoT Devices

Lawrence, Tom; Zhang, Li

doi:10.3390/s19245541

Cited by 40 publications

(28 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…11(a) shows that for CIFAR-10 the ResNet18_PSD1_P16 increases accuracy to 92% compared to the baseline MobileNetV2 (ShuffleNet) accuracy of 90.3% (∼89%). Note that our obtained accuracies are also superior than reported in [48] and only around 1% less than the accuracy reported in [49] which was trained for 180 additional epochs. The pre-defined sparse CNN model VGG16_PSD1_P8 With 0.073 G FLOPs, has approximately 1.24× (1.34×) fewer computation complexity yet still outperforms MobileNetV2 (ShuffleNet) in terms of accuracy.…”

Section: Performance Comparison With Shufflenet and Mo-bilenetv2contrasting

confidence: 56%

Pre-defined Sparsity for Low-Complexity Convolutional Neural Networks

Kundu

Nazemi

Pedram

et al. 2020

IEEE Trans. Comput.

View full text Add to dashboard Cite

The high energy cost of processing deep convolutional neural networks impedes their ubiquitous deployment in energy-constrained platforms such as embedded systems and IoT devices. This work introduces convolutional layers with pre-defined sparse 2D kernels that have support sets that repeat periodically within and across filters. Due to the efficient storage of our periodic sparse kernels, the parameter savings can translate into considerable improvements in energy efficiency due to reduced DRAM accesses, thus promising significant improvements in the trade-off between energy consumption and accuracy for both training and inference. To evaluate this approach, we performed experiments with two widely accepted datasets, CIFAR-10 and Tiny ImageNet in sparse variants of the ResNet18 and VGG16 architectures. Compared to baseline models, our proposed sparse variants require up to ∼82% fewer model parameters with 5.6× fewer FLOPs with negligible loss in accuracy for ResNet18 on CIFAR-10. For VGG16 trained on Tiny ImageNet, our approach requires 5.8× fewer FLOPs and up to ∼83.3% fewer model parameters with a drop in top-5 (top-1) accuracy of only 1.2% (∼2.1%). We also compared the performance of our proposed architectures with that of ShuffleNet and MobileNetV2. Using similar hyperparameters and FLOPs, our ResNet18 variants yield an average accuracy improvement of ∼2.8%.

show abstract

Section: Performance Comparison With Shufflenet and Mo-bilenetv2contrasting

confidence: 56%

Pre-defined Sparsity for Low-Complexity Convolutional Neural Networks

Kundu

Nazemi

Pedram

et al. 2020

IEEE Trans. Comput.

View full text Add to dashboard Cite

show abstract

“…In future directions, other hybrid leader breeding mechanisms will be explored to further enhance performance. Moreover, we also aim to evaluate the proposed models using complex computer vision tasks, e.g., deep architecture generation for object detection and classification [ 51 , 73 , 74 , 75 ] as well as image description generation [ 76 , 77 ].…”

Section: Discussionmentioning

confidence: 99%

Feature Selection Using Enhanced Particle Swarm Optimisation for Classification Models

Xie

Zhang

Lim

et al. 2021

Sensors

Self Cite

View full text Add to dashboard Cite

In this research, we propose two Particle Swarm Optimisation (PSO) variants to undertake feature selection tasks. The aim is to overcome two major shortcomings of the original PSO model, i.e., premature convergence and weak exploitation around the near optimal solutions. The first proposed PSO variant incorporates four key operations, including a modified PSO operation with rectified personal and global best signals, spiral search based local exploitation, Gaussian distribution-based swarm leader enhancement, and mirroring and mutation operations for worst solution improvement. The second proposed PSO model enhances the first one through four new strategies, i.e., an adaptive exemplar breeding mechanism incorporating multiple optimal signals, nonlinear function oriented search coefficients, exponential and scattering schemes for swarm leader, and worst solution enhancement, respectively. In comparison with a set of 15 classical and advanced search methods, the proposed models illustrate statistical superiority for discriminative feature selection for a total of 13 data sets.

show abstract

“…dermoscopic lesion localization and segmentation, nuclei counting and segmentation, MRI brain tumour segmentation, and stroke lesion segmentation based on acute CT perfusion images. We will also study the application of the proposed PSO model to other complex computer vision tasks, such as evolving deep architecture generation for image description and visual question generation [82][83][84][85] for resource-constrained deployments.…”

Section: Discussionmentioning

confidence: 99%

Intelligent optic disc segmentation using improved particle swarm optimization and evolving ensemble models

Zhang

Lim

2020

Applied Soft Computing

Self Cite

View full text Add to dashboard Cite

In this research, we propose Particle Swarm Optimization (PSO)-enhanced ensemble deep neural networks for optic disc (OD) segmentation using retinal images. An improved PSO algorithm with six search mechanisms to diversify the search process is introduced. It consists of an accelerated super-ellipse action, a refined superellipse operation, a modified PSO operation, a random leader-based search operation, an average leader-based search operation and a spherical random walk mechanism for swarm leader enhancement. Owing to the superior segmentation capabilities of Mask R-CNN, transfer learning with a PSO-based hyper-parameter identification method is employed to generate the fine-tuned segmenters for OD segmentation. Specifically, we optimize the learning parameters, which include the learning rate and momentum of the transfer learning process, using the proposed PSO algorithm. To overcome the bias of single networks, an ensemble segmentation model is constructed. It incorporates the results of distinctive base segmenters using a pixel-level majority voting mechanism to generate the final segmentation outcome. The proposed ensemble network is evaluated using the Messidor and Drions data sets and is found to significantly outperform other deep ensemble networks and hybrid ensemble clustering models that are incorporated with both the original and state-of-the-art PSO variants. Additionally, the proposed method statistically outperforms existing studies on OD segmentation and other search methods for solving diverse unimodal and multimodal benchmark optimization functions and the detection of Diabetic Macular Edema.

show abstract

IoTNet: An Efficient and Accurate Convolutional Neural Network for IoT Devices

Cited by 40 publications

References 38 publications

Pre-defined Sparsity for Low-Complexity Convolutional Neural Networks

Pre-defined Sparsity for Low-Complexity Convolutional Neural Networks

Feature Selection Using Enhanced Particle Swarm Optimisation for Classification Models

Intelligent optic disc segmentation using improved particle swarm optimization and evolving ensemble models

Contact Info

Product

Resources

About