Neural Network Data Processing Apparatus with Segmented Computing Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) applied to embedded systems face efficiency issues due to the large number of product-sum operations and varying feature image resolutions, leading to decreased processing efficiency and increased processing time.
Innovation Solution
A data processing apparatus with separate units for handling different types of data, including a first product-sum operation unit, a branch unit for parallel output, a sampling unit, and a second product-sum operation unit, which allows for efficient processing of feature images with varying resolutions by optimizing the number of operations and parallel processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a common computing unit is used to perform product-sum operations on each layer of the CNN, then the cost of the data processing apparatus is suppressed, but the processing efficiency deteriorates when feature images have different resolutions
Solution Approach 1:
The patent divides the computing unit into multiple specialized computing units, each dedicated to processing feature images of a specific resolution. This segmentation allows each unit to operate at full efficiency for its designated resolution while maintaining cost-effectiveness through specialized optimization, resolving the contradiction between apparatus cost and processing efficiency.
2Adaptability or versatility
If sampling processing is used to match the sizes of feature images with different resolutions, then the network configuration can be maintained, but the processing efficiency of the computing unit deteriorates
Solution Approach 1:
Instead of using sampling processing that reduces efficiency, the patent segments the computing unit into multiple units handling different resolutions. This allows the network to maintain its flexible configuration with multiple resolutions while each computing unit processes features at its native resolution without inefficient sampling operations.
3Adaptability or versatility
If the number of feature images does not correspond to the number of feature images that can be processed in parallel by the computing unit, then the network structure is maintained, but the processing efficiency decreases
Solution Approach 1:
The patent segments the computing unit into multiple specialized units, each optimized for specific resolutions. This segmentation allows the system to maintain flexible network structures with varying numbers of feature images while each computing unit operates at full parallel processing efficiency for its designated resolution, eliminating the efficiency loss that would occur with a single computing unit.
Data Source
AI summary
An apparatus includes a first holding unit and a second holding unit configured to hold first-type data and second-type data, respectively, a first operation unit configured to execute a first product-sum operation based on the first-type data, a branch unit configured to output an operation result of the first product-sum operation in parallel, a sampling unit configured to sample the operation result and to output a sampling result, and a second operation unit configured to execute a second product-sum operation based on the second-type data and the sampling result.