Neural Network Integral Map Generation for Low-Power Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network devices require significant energy consumption due to frequent memory access for processing large amounts of data, which is a challenge for low-power devices like mobile and IoT systems.
Innovation Solution
A method and apparatus that generate an integral map for each channel in a neural network's first layer by calculating area sums of pixel values, perform a convolution operation to obtain pre-output feature maps of a subsequent layer, and remove offsets in weight kernels by subtracting accumulated values from pixel values, using asymmetric quantization weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If large amounts of data are processed through frequent memory access, then computation capability is improved, but energy consumption increases
Solution Approach 1:
The patent pre-calculates and stores integral maps (cumulative sums of pixel values) before the main convolution operation. This preliminary action allows the system to quickly compute sum of products without repeatedly accessing memory during the computation, thereby improving computation capability while reducing energy consumption from frequent memory accesses
Solution Approach 2:
The patent transforms the computation from directly processing pixel values to processing integral maps, which are pre-computed cumulative sums. This dimensional transformation allows the system to perform convolutions more efficiently by working with pre-aggregated data, reducing the need for frequent memory access during computation
2Loss of energy
If asymmetric quantization is used to reduce computation complexity, then energy consumption is reduced, but precision of weight kernels may be affected
Solution Approach 1:
The patent applies asymmetric quantization to transform weight kernels from floating-point to integer representation, changing the parameter format to reduce computation complexity and energy consumption. The asymmetric nature allows better representation of the actual weight distribution, maintaining precision while enabling efficient integer arithmetic operations
Solution Approach 2:
The patent uses integral maps as a copy of the original pixel data, pre-computed and stored for efficient access. This copying approach allows the system to work with pre-aggregated values instead of repeatedly accessing and computing from original pixel data, reducing memory access frequency and energy consumption
Data Source
AI summary
A method of processing a neural network, includes generating an integral map for each channel in a first layer of the neural network based on calculating of area sums of pixel values in first output feature maps of channels in the first layer, generating an accumulated integral map by performing an accumulation operation on the integral maps generated for the respective channels, obtaining pre-output feature maps of a second layer, subsequent to the first layer, by performing a convolution operation between input feature maps of the second layer and weight kernels, and removing offsets in the weight kernels to obtain second output feature maps of the second layer by subtracting accumulated values of the accumulated integral map from pixel values of the pre-output feature maps.