Merging Neural Network Operation Groups to Reduce Redundancy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems face inefficiencies due to redundant computation of identical operation groups across different neural networks, leading to increased memory use, power consumption, and processing time.
Innovation Solution
A method is introduced to merge identical operation groups from multiple neural networks into a shared operation group, which is then computed once using a selected hardware resource, reducing redundant execution and optimizing resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If identical operation groups in multiple neural networks are executed separately, then each neural network can be processed independently, but redundant computations increase processing time and energy consumption
Solution Approach 1:
The patent merges identical operation groups from multiple neural networks into a single shared operation group that is executed once by a processor. This combining of redundant computational tasks eliminates repeated processing of the same operations across different neural networks, thereby reducing energy consumption while maintaining processing efficiency.
2Device complexity
If identical operation groups are executed separately in each neural network, then implementation is simpler, but memory usage increases due to redundant storage
Solution Approach 1:
The patent combines multiple identical operation groups into a single shared operation group that is stored in memory once and referenced by multiple neural networks. This merging approach reduces the total quantity of operational data stored in memory while maintaining system functionality, thus reducing memory usage without significantly increasing system complexity.
3Reliability
If separate processing of operation groups is used, then neural network independence is maintained, but overall system performance decreases due to redundant computations
Solution Approach 1:
The patent merges identical operation groups into a shared structure that can be efficiently executed once and reused across multiple neural networks. This approach improves overall system performance by eliminating redundant computations while maintaining the logical independence of each neural network through shared operational references.
Data Source
AI summary
A method of operating a neural network system includes merging, by a processor, a first operation group in a first neural network and a second operation group in a second neural network, including identical operations, as a shared operation group; selecting, by the processor, a first hardware to execute the shared operation group, from among a plurality of hardware; and executing the shared operation group by using the first hardware.