CNN Building Blocks with Channel Split and Shuffle Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require extensive computational resources and energy consumption, and the use of different types of convolutional units complicates their implementation, necessitating an efficient solution.
Innovation Solution
The proposed solution involves a neural network processor with multiple building blocks, each comprising a channel split unit, a convolution unit, a concatenation unit, and a shuffle unit, which work together to perform efficient convolution operations and reduce computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional CNN architectures are used, then computational accuracy is maintained, but computational resources and energy consumption increase significantly
Solution Approach 1:
The network is divided into multiple building blocks, each containing specific convolutional layers with fixed kernel sizes (3x3, 5x5, 7x7). This segmentation allows for optimized resource allocation and reduced computational overhead in each module while maintaining overall network performance.
Solution Approach 2:
The patent changes the parameter of convolutional kernel sizes to fixed values (3x3, 5x5, 7x7) rather than using variable or larger kernels. This parameter optimization reduces the number of computations required while preserving the essential feature extraction capabilities of the CNN.
2Adaptability or versatility
If different types of convolutional units are used to improve network capability, then processing power increases, but implementation complexity increases
Solution Approach 1:
The building block design serves multiple functions: it contains convolutional layers with different kernel sizes, batch normalization, activation functions, and pooling operations all within a single standardized module. This multi-functionality reduces implementation complexity while maintaining network capability.
Solution Approach 2:
By segmenting the network into identical building blocks, the patent simplifies implementation despite incorporating diverse operations. Each block is a self-contained unit that can be independently implemented and reused, reducing overall system complexity.
Data Source
AI summary
An apparatus that may include a neural network processor, the neural network processor comprises multiple building blocks. Each of the at least some of the building blocks may include, may consist or may consist essentially of a channel split unit, a convolution unit, a concatenation unit, and a shuffle unit.