Convolutional Neural Network Quantization via Layer-Specific Fixed-Point Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks face challenges in computation speed and accuracy due to high computational and storage demands, necessitating improvements in these areas.
Innovation Solution
A computation method and device that determine optimal quantization step sizes based on data distribution, applying fixed-point processing to improve calculation efficiency and reduce storage requirements by iteratively optimizing quantization step sizes across layers of the neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point arithmetic is used in convolutional neural networks, then computational accuracy is maintained, but computation speed decreases and storage space increases
Solution Approach 1:
The patent applies parameter changes by transitioning from floating-point to fixed-point arithmetic, changing the numerical representation parameters to achieve faster computation and reduced storage while maintaining acceptable accuracy through optimized quantization step sizes
Solution Approach 2:
The patent implements dynamics by dynamically determining optimal quantization step sizes for different layers based on data distribution characteristics, allowing the system to adapt the fixed-point precision requirements to actual computational needs
2Measurement precision
If floating-point arithmetic is used in convolutional neural networks, then computational accuracy is maintained, but storage space requirements increase
Solution Approach 1:
The patent changes the data representation parameter from floating-point to fixed-point format, reducing the number of bits required for storage while maintaining computational accuracy through carefully selected quantization step sizes that preserve essential data characteristics
3Quantity of substance
If fixed-point processing is applied to all layers with uniform precision, then storage space is reduced, but computational accuracy deteriorates
Solution Approach 1:
The patent applies local quality by determining optimal quantization step sizes specific to each layer based on the data distribution characteristics of that layer, allowing different layers to use appropriate precision levels rather than uniform precision throughout
Solution Approach 2:
The patent implements dynamics by adaptively determining quantization step sizes for each layer based on actual data distributions, allowing the system to optimize the balance between storage efficiency and computational accuracy for each specific layer
Data Source
AI summary
A computation method used in a convolutional neural network is provided. The method includes: receiving original data; determining a first optimal quantization step size according to a distribution of the original data; performing fixed-point processing to the original data according to the first optimal quantization step size to generate first data; inputting the first data to a first layer of the convolutional neural network to generate first output data; determining a second optimal quantization step size according to a distribution of the first output data; performing the fixed-point processing to the first output data according to the second optimal quantization step size to generate second data; and inputting the second data to a second layer of the convolutional neural network.