DNN Layer Operation Using Fixed-Point Kernel Splitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Deep Neural Networks (DNNs) suffer from low operating efficiency due to the large amount of data involved in operations, leading to a significant number of operations in each task.
Innovation Solution
The proposed solution involves an operation method and apparatus for a network layer in a DNN, where a weighted tensor is acquired and each filter is split into a linear combination of fixed-point convolution kernels with specified bit-width weights, reducing the number of operations by performing fixed-point quantization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If floating-point multiplication/addition operations are used in conventional DNN, then calculation precision is maintained, but operating efficiency is low due to large amount of data and operations
Solution Approach 1:
The patent changes the numerical representation parameter from floating-point to fixed-point format. By quantizing weights to fixed-point values with specified bit-width, the operations can be performed more efficiently with reduced computational complexity, directly improving operating efficiency and reducing operation time
2Productivity
If filter is used as single unit for convolution operation, then operation structure is simple, but amount of operations is large leading to low efficiency
Solution Approach 1:
The patent segments the filter into multiple sub-filters based on weight quantization levels. Each sub-filter corresponds to a specific bit-width group, allowing the convolution operation to be divided into multiple smaller operations. This segmentation reduces the computational burden of each individual operation while maintaining the overall filtering capability, thereby reducing the total amount of operations required
Data Source
AI summary
An operation method and apparatus for a network layer in a Deep Neural Network are provided. The method includes: acquiring a weighted tensor of the network layer in the Deep Neural Network, the weighted tensor comprising a plurality of filters; converting each filter into a linear combination of a plurality of fixed-point convolution kernels by splitting the filter, wherein a weight value of each of the fixed-point convolution kernels is a fixed-point quantized value having a specified bit-width; for each filter, performing a convolution operation on input data of the network layer and each of the fixed-point convolution kernels, respectively, to obtain a plurality of convolution results, and calculating a weighted sum of the obtained convolution results based on the linear combination of the plurality of fixed-point convolution kernels of the filter to obtain an operation result of the filter; determining output data of the network layer.