CNN Inference With Separate Target And Margin Data for Lower SRAM Use
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of reducing the storage capacity required for convolutional neural network computations while maintaining computation result reliability is addressed by the present disclosure, which involves obtaining target data and margin data separately and using padding to minimize the SRAM capacity without compromising accuracy.
Innovation Solution
An information processing apparatus that executes convolutional computations using a convolutional neural network by obtaining target data and margin data separately, with a portion of the margin data being obtained from the original image and the rest padded with zeros to reduce SRAM capacity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If zero padding is performed to reduce SRAM capacity, then storage capacity is reduced, but reliability of computation result declines
Solution Approach 1:
The patent segments the filter into multiple sub-filters and divides the input image into multiple patches. Each patch is processed by corresponding sub-filters, and results are aggregated. This segmentation allows selective application of zero padding only to specific boundary regions rather than the entire filter, thereby reducing SRAM capacity while maintaining computation reliability in non-padded regions.
Solution Approach 2:
The patent applies zero padding selectively to specific local regions (boundary areas) of the input image rather than uniformly across the entire image. This local quality approach ensures that only necessary boundary regions are padded with zeros, preserving the reliability of computation results in non-padded regions while minimizing the overall SRAM capacity required for storing padded data.
2Reliability
If hardware resource is increased to maintain computation reliability, then reliability is enhanced, but device cost and size increase
Solution Approach 1:
The patent segments the convolution operation into multiple smaller operations using sub-filters and image patches. This segmentation reduces the hardware resource requirements for each individual operation, allowing the system to maintain computation reliability through multiple passes rather than requiring large-scale hardware resources in a single operation.
Solution Approach 2:
The patent applies zero padding partially rather than excessively - only to the boundary regions where it is strictly necessary for maintaining computation reliability. This partial action approach minimizes the hardware resources required for padding operations while still ensuring reliable computation results in the non-padded regions.
3Measurement precision
If filter size is increased to improve pattern recognition, then recognition accuracy is improved, but SRAM capacity requirement increases
Solution Approach 1:
The patent segments a large filter into multiple smaller sub-filters, each operating on specific patches of the input image. This segmentation maintains the overall pattern recognition capability of the large filter while reducing the SRAM capacity required for each individual sub-filter operation, as smaller filters require less storage space.
Solution Approach 2:
The patent applies zero padding partially to boundary regions rather than expanding the filter size uniformly. This allows the effective filter size to be maintained for pattern recognition accuracy in non-boundary regions, while minimizing the SRAM capacity increase that would result from uniformly expanding the filter across the entire image.
Data Source
AI summary
An information processing apparatus configured to execute inference using a convolutional neural network, including: an obtainment unit configured to obtain target data from data for inference inputted in the information processing apparatus; and a computation unit configured to execute convolutional computation and output computation result data, the convolutional computation using computation data including the target data obtained by the obtainment unit and margin data different from the target data that is required to obtain the computation result data in a predetermined size, in which the obtainment unit obtains first data, which is a part of the margin data, from a data group existing around the target data separately from the target data in the data for inference and doses not obtain second data, which is the margin data except the first data, from the data group.