Neural Network Architecture Search Optimizing Processor Computation Cost
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining neural network architectures do not effectively guarantee high-speed execution and reduced power consumption on processors, as they fail to optimize computation costs and prediction errors during the training process.
Innovation Solution
A method and apparatus that search for a target neural network architecture based on a processor computation cost, which includes time-consuming and power-consuming hyperparameters, by repeatedly searching for and evaluating neural network architectures until specific end conditions are met, ensuring optimal performance and efficiency during model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If existing methods for determining neural network architectures are used, then the architecture can be obtained, but high-speed execution and reduced power consumption on processors cannot be guaranteed
Solution Approach 1:
The patent changes the optimization parameters from traditional accuracy-only metrics to include processor computation cost, time-consuming hyperparameters, and power-consuming hyperparameters. By modifying the loss function to incorporate these processor-specific parameters, the neural network architecture is optimized for both speed and reliability simultaneously, resolving the contradiction between execution speed and performance guarantee.
2Use of energy by moving object
If existing neural network architecture determination methods are used, then the model can be trained, but computation costs and power consumption are not optimized
Solution Approach 1:
The patent introduces power-consuming hyperparameters and time-consuming hyperparameters as additional optimization parameters in the loss function. This allows the architecture search process to simultaneously optimize for power consumption and training efficiency, rather than treating them as conflicting objectives. The modified loss function guides the search toward architectures that achieve both low power consumption and high training efficiency.
3Adaptability or versatility
If neural network architecture search is performed without considering processor computation cost, then various architectures can be explored, but optimal performance and efficiency cannot be ensured
Solution Approach 1:
The patent incorporates processor computation cost as a specific parameter in the loss function, which allows the architecture search to maintain flexibility in exploring various neural network structures while simultaneously optimizing for processor-specific performance metrics. The computation cost parameter acts as a constraint that guides the search toward architectures that are both versatile and precisely optimized for the target processor.
Data Source
AI summary
A method and apparatus for determining a neural network architecture of a processor are provided. The method of determining a target neural network architecture, the method comprising obtaining a first neural network architecture, searching for the first neural network architecture based on a loss function, in response to a first search end condition not being satisfied, and determining a target neural network architecture used in a processor, based on a result of the searching, wherein the loss function is based on a processor computation cost.