Neural Network Compression Through Calculator-Aware Operation Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks used in vehicle operation devices face challenges in reducing processing time and optimizing the use of multiple calculators, leading to inefficiencies in calculation processing.
Innovation Solution
A neural network reduction device that allocates calculators for calculation processing, sets the number of operations based on allocated calculators, and reduces the neural network to match the set operations, using techniques like pruning to minimize unnecessary operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the number of operations of the neural network is reduced by reducing branches, then the calculation load is reduced, but the processing time is not necessarily reduced and multiple calculators are not effectively used
Solution Approach 1:
The patent segments the neural network operations into distinct calculation graphs that can be independently allocated to multiple calculators. By dividing the computation into separable tasks with specific operation counts, the system can distribute work across multiple calculators efficiently, ensuring that reducing the number of operations actually reduces processing time when multiple calculators are deployed.
Solution Approach 2:
The patent introduces dynamic allocation of calculators to calculation graphs based on the number of operations required. The system can adjust the number of calculators allocated to each graph dynamically, allowing optimal resource utilization. This dynamic approach ensures that the processing time is reduced effectively by matching the computational workload with the available calculator resources.
2Device complexity
If the number of operations is reduced without considering calculator allocation, then the neural network is simplified, but multiple calculators are not effectively utilized
Solution Approach 1:
The patent segments the neural network into multiple calculation graphs with different numbers of operations, enabling fine-grained control over resource allocation. This segmentation allows the system to match the complexity of each calculation graph with an appropriate number of calculators, ensuring effective utilization while maintaining network simplification.
Solution Approach 2:
The patent changes the parameter of operation count for each calculation graph to be a multiple of the allocated number of calculators. By adjusting this parameter, the system ensures that all calculators can be effectively utilized without idle resources, while still achieving the goal of reducing overall network complexity.
3Productivity
If branches with small weighting factors are removed, then the calculation load is reduced, but the processing time may not decrease and calculator efficiency drops
Solution Approach 1:
The patent segments the removal of branches into controlled operations where the resulting calculation graphs are designed to have operation counts compatible with the number of available calculators. This segmentation ensures that the simplification process maintains calculator efficiency by creating balanced workload distributions.
Solution Approach 2:
The patent applies parameter changes by ensuring that the number of operations in each calculation graph is adjusted to be a multiple of the allocated calculator count. This parameter adjustment maintains processor efficiency while achieving the reduction in calculation load through branch removal.
Data Source
AI summary
Processing time of a neural network is shortened, and the number of operations of the neural network is reduced such that a plurality of calculators can be effectively used. A neural network reduction device (100) that reduces the number of operations of a neural network by an operation device (140) including a plurality of calculators by reducing the neural network, the neural network reduction device including: a calculator allocation unit (102) that sets the number of calculators allocated to calculation processing of the neural network; a number-of-operations setting unit (103) that sets the number of operations of a reduced neural network based on the number of allocated calculators; and a neural network reduction unit (104) that reduces the neural network such that the number of operations of the neural network by the operation device (140) is equal to the number of operations set by the number-of-operations setting unit (103).