CNN Input Buffer Loading for Real-Time Embedded Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Sophisticated convolutional neural networks (CNNs) require significant resources and are often not operable in real time on less capable terminals like smartphones due to their computational demands.
Innovation Solution
A processor-implemented method for CNN processing that involves determining a loading space unit based on input and kernel dimensions, interleaving input elements in memory, and performing selective convolution operations using kernel information to optimize resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sophisticated CNNs are implemented to improve recognition performance, then accuracy is improved, but resource consumption increases making them inoperable on lesser capable terminals
Solution Approach 1:
The patent segments the CNN processing into multiple stages: loading only necessary input elements into buffer memory, performing selective convolution operations, and using skip connections to bypass unnecessary computations. This segmentation allows sophisticated CNNs to run on resource-constrained devices by dividing the computational task into manageable parts that can be executed with limited resources.
Solution Approach 2:
The patent implements partial action by loading only the necessary portion of input elements into buffer memory rather than the entire input dataset. The convolution operations are performed selectively on these loaded elements, and skip connections allow the network to bypass certain computational paths when appropriate, reducing overall resource consumption while maintaining recognition accuracy.
2Measurement precision
If sophisticated CNNs are implemented to improve recognition performance, then accuracy is improved, but real-time operation becomes infeasible due to computational demands
Solution Approach 1:
The patent applies preliminary action by pre-loading necessary input elements into buffer memory before convolution operations are performed. This pre-positioning of data in fast-access memory reduces computation time during actual processing, enabling real-time operation while maintaining the accuracy of sophisticated CNNs.
Solution Approach 2:
By segmenting the CNN into modular components with skip connections, the patent enables parallel and selective execution of computational tasks. This segmentation allows the system to process information more efficiently and respond in real-time while maintaining high recognition accuracy through the preserved architectural complexity.
3Speed
If more resources are allocated to CNN processing, then processing speed is improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent implements self-service through skip connections that allow the CNN architecture to automatically bypass unnecessary computational paths based on the input data characteristics. This self-optimizing mechanism improves processing speed without requiring external resource allocation or increasing device complexity, as the system adapts its own resource usage dynamically.
Data Source
AI summary
Disclosed is a convolutional neural network (CNN) processing apparatus and method, the apparatus configured to determine a loading space unit for at least one loading space in an input based on a height or a width for an input feature map of the input and an extent of a dimension of a kernel feature map, load target input elements corresponding to a target loading space, among the at least one loading space, from a memory and store the target input elements in an allocated input buffer having a size corresponding to the loading space unit, and perform a convolution operation between the target input elements stored in the input buffer and at least one kernel element of a kernel.