Neural Network Construction Device for Embedded Hardware Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for designing neural networks are time-consuming due to the need for repeated trial-and-error tests to find an optimal network that meets hardware constraints, particularly in embedded devices, which hinders efficient development and deployment.
Innovation Solution
A neural network construction device and method that narrows down candidate hyperparameters by setting scale and performance constraints based on resource information, generating models that meet these constraints, and evaluating their accuracy, thereby reducing the need for extensive testing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If trial-and-error tests are performed to determine optimal neural network for hardware constraints, then the neural network meets the hardware constraints, but the development time increases significantly
Solution Approach 1:
The patent applies preliminary action by calculating the scale constraint before neural network construction. The scale constraint is computed based on hardware information (memory size, CPU frequency) and problem information (input/output dimensionality) prior to model generation. This preliminary calculation filters out infeasible hyperparameter combinations beforehand, preventing time-wasting trial-and-error tests during the development phase.
Solution Approach 2:
The patent segments the constraint satisfaction process into two distinct phases: (1) scale constraint calculation based on hardware and problem information, and (2) neural network construction using only hyperparameters that satisfy this constraint. This segmentation allows the system to separately handle feasibility checking and model optimization, improving overall efficiency.
2Measurement precision
If extensive trial-and-error testing is performed to evaluate neural network candidates, then the optimal model is found, but computational resources are consumed excessively
Solution Approach 1:
The patent performs preliminary filtering by calculating the scale constraint before evaluating neural network candidates. This preliminary action eliminates infeasible models that would not meet hardware constraints, reducing the number of candidates requiring extensive computational evaluation and thus lowering resource consumption.
Solution Approach 2:
The patent extracts and applies the scale constraint as a filtering mechanism before the main evaluation process. By taking out the feasibility check and performing it separately beforehand, the system avoids wasting computational resources on evaluating models that cannot be deployed on the target hardware.
3Productivity
If the scale of neural network is increased to improve performance, then the processing performance improves, but the hardware resource requirements increase
Solution Approach 1:
The patent uses parameter changes by calculating the scale constraint as a function of hardware parameters (memory size, CPU frequency) and problem parameters (input/output dimensionality). This mathematical relationship allows the system to determine the maximum feasible network scale that balances performance requirements with hardware resource limitations.
Solution Approach 2:
The patent performs preliminary calculation of the scale constraint to establish the maximum feasible network size before model construction. This preliminary determination ensures that the neural network is designed within hardware limits from the outset, avoiding the need for later scaling adjustments that would require additional resources.
Data Source
AI summary
A neural network construction device includes: an obtainer which obtains resource information related to a computational resource of an embedded device and performance constraints related to processing performance of the embedded device; a setting unit which sets scale constraints of a neural network on the basis of the resource information; a generator which generates a model of the neural network on the basis of the scale constraints; and a determination unit which determines whether or not the model generated meets the performance constraints, and outputs data based on the result of the determination.