Scholarlink

Convolutional Neural Networks (CNN) are composed of hundreds of millions of dot operations, leading to the emergence of new algorithms to reduce the high computational complexity. This paper introduces input repetition: a new solution to reduce the number of operations of previous quantized CNN models without degrading the accuracy. The input repetition occurs when several identical elements are found in a group of input activations, allowing the reduction of the number of multipliers required to perform the dot operation needed in the convolutional layers. This paper analyzes several strategies to find repetition in the convolution input using quantized state-of-the-art CNN: YoloV3, MobileNET and ResNet50. Results show that input repetition can reduce multiplications by a factor of 2.4x without any performance impact.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

On the Potentials of Input Repetition in CNN Networks for Reducing Multiplications

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Laura Medina,
  • Jose Flich

摘要

Convolutional Neural Networks (CNN) are composed of hundreds of millions of dot operations, leading to the emergence of new algorithms to reduce the high computational complexity. This paper introduces input repetition: a new solution to reduce the number of operations of previous quantized CNN models without degrading the accuracy. The input repetition occurs when several identical elements are found in a group of input activations, allowing the reduction of the number of multipliers required to perform the dot operation needed in the convolutional layers. This paper analyzes several strategies to find repetition in the convolution input using quantized state-of-the-art CNN: YoloV3, MobileNET and ResNet50. Results show that input repetition can reduce multiplications by a factor of 2.4x without any performance impact.