CNN Convolution With Selective Margin Data for Reduced SRAM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require significant hardware resources due to massive computations, and reducing SRAM capacity through zero padding leads to unreliable computation results, making implementation in products challenging.
Innovation Solution
An information processing apparatus that obtains target data and margin data separately, using a data obtainment condition to determine convolutional computation results, reducing SRAM capacity while maintaining reliability.
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 applies local quality by differentiating between peripheral regions (where zero padding is applied) and central regions (where actual image data is retained). This allows the system to reduce overall SRAM capacity while preserving the reliability of computation results in critical areas. The convolution filter selectively processes valid image data in central regions while handling padded regions appropriately, thus maintaining computational reliability despite reduced storage capacity.
2Reliability
If hardware resource is increased to enhance reliability of computation result, then reliability is improved, but device complexity and cost increase
Solution Approach 1:
The patent segments the image processing task into regions requiring full precision (central regions with actual image data) and regions where approximation is acceptable (peripheral regions with zero padding). This segmentation allows the system to achieve reliable computation results for critical features while reducing overall hardware resource requirements. The convolution operation is performed selectively across different regions with appropriate data handling strategies.
3Ease of manufacture
If SRAM capacity is reduced for product implementation, then ease of manufacture is improved, but reliability of computation result is compromised
Solution Approach 1:
The patent changes the parameter of data representation by introducing zero padding in peripheral regions while maintaining actual image data in central regions. This parameter change enables the system to operate with reduced SRAM capacity (improving ease of manufacture and product implementation) while preserving computation reliability through selective data retention and appropriate convolution processing that accounts for padded regions.
Data Source
AI summary
An information processing apparatus includes: 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, according to a data obtainment condition set in advance, the obtainment unit obtains first data, which is a part of the margin data, from a data group existing around 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, and a coefficient of the convolutional computation is determined by learning to which the data obtainment condition is reflected.