SSP Framework: A New Approach to Designing Lightweight Convolutional Neural Networks

Creative Commons License

Turay T., Vladimirova T.

IEEE ACCESS, cilt.10, ss.102292-102307, 2022 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 10
  • Basım Tarihi: 2022
  • Doi Numarası: 10.1109/access.2022.3208924
  • Dergi Adı: IEEE ACCESS
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.102292-102307
  • Anahtar Kelimeler: Computational modeling, Convolutional neural networks, Computer vision, Memory management, System-on-chip, Object detection, Field programmable gate arrays, Embedded systems, Image classification, Autonomous driving, Computer Vision, convolutional neural networks, convolution types, real-time embedded systems, object detection, image classification, visual recognition, autonomous driving systems, CONNECTIONS, CNN
  • Atatürk Üniversitesi Adresli: Hayır

Özet

Most convolutional neural network (CNN) designs are still bottlenecked by costly computational load, which impedes their utilization in industrial applications. To address this issue, a new Sparse-Split-Parallelism (SSP) design framework is proposed in this paper. It fuses three design strategies that can be applied to the majority of the popular state-of-the-art block-based CNN models to lighten their computing budget while maintaining comparable accuracies. At a block level, a design strategy, based on the novel concept of sparse skip connections, is introduced, which provides optimal connectivity, preventing a severe rise in channel numbers and keeping a satisfactory feature reuse in the network model. As part of the module-level design, a new SSP module is created that preserves the design features of the targeted existing models, and a novel proportional channel split operation is employed to achieve optimal trade-off between accuracy and model size. As the third strategy at a layer level, the idea of the degree of parallelism is adopted, resulting into an equal number of channels in the layers, which decreases memory access and yields a better inference time. The effectiveness of the framework has been validated through a comprehensive experimental work. The evaluation results, which are based on DenseNet, ResNet, ShiftNet, ShuffleNet, and ShuffleNet-v2, verify that the proposed SSP framework is notably capable of reducing the parameter number, FLOPs, and inference time of existing CNN models with quite alike accuracies. The models are evaluated on image classification using the ImageNet and CIFAR-10&CIFAR-100 datasets, as well as on object detection with the MS COCO dataset.