CNN Convolutional Layer Using Fixed Multipliers and Adders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require significant storage resources and computing power due to the varying sizes of convolution kernels in different layers, leading to complex storage control and high power consumption.
Innovation Solution
The method involves using a fixed number of multipliers and adders for each convolutional layer to perform convolution operations, allowing for efficient multiplication and addition operations across all kernels in a single cycle, reducing the need for complex storage control and intermediate buffering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If convolution kernels of different sizes are used in different convolutional layers, then the CNN can process various feature patterns effectively, but storage control complexity and power consumption increase
Solution Approach 1:
The patent applies universality by designing a unified convolutional layer structure that can handle multiple kernel sizes (10×1×1, 10×1×16, 6×1×16) through a single standardized interface. The convolutional layer uses consistent connection patterns to input and output data regardless of the specific kernel size being applied, allowing the same hardware structure to perform multiple functions with different kernel dimensions without requiring separate specialized circuits for each kernel size.
2Adaptability or versatility
If convolution kernels of different sizes are used in different convolutional layers, then the CNN can process various feature patterns effectively, but power consumption increases
Solution Approach 1:
The patent reduces power consumption by creating a universal convolutional layer design that reuses the same computational units and data pathways for all kernel sizes. Instead of implementing separate dedicated circuits for each kernel dimension, the system uses a single configurable structure that adapts to different kernel sizes, minimizing the activation of additional hardware components and reducing overall power consumption while maintaining the ability to process diverse feature patterns.
3Adaptability or versatility
If varying kernel sizes are implemented across convolutional layers, then diverse feature processing is enabled, but intermediate buffering requirements increase
Solution Approach 1:
The patent minimizes intermediate buffering requirements by designing a unified data flow architecture where the same input and output data structures are used across all convolutional layers regardless of kernel size. The standardized connection patterns allow data to flow through different kernel processing units without requiring size-specific intermediate buffers, as the universal structure maintains consistent data dimensions and formats throughout the network.
Data Source
AI summary
A method for improving a convolutional neural network (CNN) to perform computations is provided. The method includes the following steps: determining a number of a plurality of multipliers to be N and a number of a plurality of adders to be N according to a number of convolution kernels used by a plurality of convolution layers; and in response to an i-th convolutional layer of the convolutional neural network performing a convolution operation and N convolution kernels of the i-th convolutional layer being all in a size of K×1×1, using the N multipliers and the N adders to perform a multiplication operation once and an addition operation once for each of the N convolution kernels of the i-th convolutional layer in one cycle, such that N outputs of the N convolution kernels of the i-th convolutional layer are obtained after K cycles.