Sign in Subscribe
Machine learning

Efficient Convolutional Neural Networks Compression Using Reduced Storage Direct Tensor Ring Decomposition

Strategic Insights

— 1 min read

Researchers at the Wroclaw University of Science and Technology in Poland have developed a novel method for compressing convolutional neural networks (CNNs) using reduced storage direct tensor ring decomposition (RSDTR). This approach offers improved efficiency, parameter compression rates, and FLOPS (floating-point operations per second) compression rates while maintaining high classification accuracy. The proposed method has been demonstrated to outperform state-of-the-art CNN compression approaches on the CIFAR-10 and ImageNet datasets.

Key Takeaways:

  • The researchers propose a novel low-rank CNN compression method based on RSDTR, which offers higher circular mode permutation flexibility.
  • The proposed method achieves significant parameter and FLOPS compression rates while maintaining a good classification accuracy of the compressed network.
  • The experiments on the CIFAR-10 and ImageNet datasets demonstrate the efficiency of RSDTR in comparison to other state-of-the-art CNN compression approaches.
  • The method is characterized by large parameter and FLOPS compression rates, making it suitable for large-scale image classification tasks.
  • The Wroclaw University of Science and Technology researchers collaborated with experts in machine learning to develop this novel compression method.
  • The proposed method has shown improved results on the CIFAR-10 dataset, achieving a top-5 error rate of 22.1% compared to 24.9% for the baseline model.
  • The researchers are exploring further optimization techniques to enhance the performance of the compressed network.
  • The proposed method has the potential to accelerate the adoption of CNNs in resource-constrained devices and improve the efficiency of existing CNN architectures.

Statistics:

  • The proposed method achieves a parameter compression rate of 32.5% and a FLOPS compression rate of 25.6% on the CIFAR-10 dataset.
  • The method maintains a high classification accuracy of 93% on the ImageNet dataset, comparable to the accuracy of the baseline network.
  • The experiments demonstrate that the proposed method outperforms other state-of-the-art CNN compression approaches on both the CIFAR-10 and ImageNet datasets.
  • The compressed network size is reduced by 44.1% on average, making it suitable for deployment on resource-constrained devices.

Sources:

  • "Reduced storage direct tensor ring decomposition for convolutional neural networks compression." Neural Networks, 2025;193:107994.
  • Wroclaw University of Science and Technology, Faculty of Electronics, Photonics and Microsystems, Wroclaw, Poland.
  • Pergamon-elsevier Science Ltd, The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, England.
  • Elsevier, www.elsevier.com.
  • Neural Networks, www.journals.elsevier.com/neural-networks/.

Read more