Neural Network Architecture Search via Aging Evolution Algorithm
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manually determining the optimal number of residual blocks, skip connections, and other parameters for a neural network to perform well in keyword spotting tasks is burdensome and inefficient.
Innovation Solution
The use of an aging evolution algorithm for neural network architecture search, which involves mutating neural network architectures by adding or removing residual blocks, modifying parameters such as skip connections, and evaluating these architectures using a performance metric like keyword spotting accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual methods are used to determine neural network parameters, then design control is maintained, but the process becomes burdensome and inefficient
Solution Approach 1:
The system performs self-service by automatically searching for optimal neural network architectures through evolutionary algorithms. The aging evolution algorithm autonomously mutates, evaluates, and selects architectures without human intervention, transforming the manual design process into an automated self-optimizing system that improves productivity while reducing operational burden
Solution Approach 2:
The patent replaces the mechanical manual design process with an automated computational system. The evolutionary algorithm substitutes human expertise and manual tuning with algorithmic search mechanisms that automatically explore the architecture space, evaluating and selecting optimal configurations based on performance metrics
2Reliability
If the number of residual blocks and skip connections is increased, then network performance improves, but computational resources and complexity increase
Solution Approach 1:
The system dynamically changes architectural parameters including the number of residual blocks, skip connections, and other network configurations through evolutionary mutation. The aging evolution algorithm explores different parameter combinations and selects those that optimize keyword spotting accuracy while managing computational complexity, finding the optimal balance between performance and resource requirements
Solution Approach 2:
The patent introduces dynamics by making the network architecture adaptable and evolvable. Rather than using fixed architectures, the system continuously mutates and evolves network configurations, allowing the architecture to dynamically adjust the number of residual blocks and skip connections based on performance feedback, achieving optimal complexity-performance trade-offs
3Measurement precision
If extensive architecture search is performed, then optimal performance is achieved, but computational resources and time are consumed
Solution Approach 1:
The aging evolution algorithm employs periodic action by systematically evaluating architectures at regular intervals and implementing aging mechanisms where older architectures are replaced by newer, potentially better ones. This periodic evaluation and replacement cycle enables thorough exploration of the architecture space while managing search time through structured, rhythmic assessment intervals
Solution Approach 2:
The system performs preliminary action by pre-evaluating and filtering architectures during the search process, using performance metrics to identify promising candidates early. The aging mechanism prepares and maintains a population of architectures, performing preliminary assessments before full evaluation, thus reducing the overall time required to find optimal solutions
Data Source
AI summary
A system and a method are disclosed for neural network architecture search. In some embodiments, the method includes: performing a neural network architecture search, wherein: the performing of the neural network architecture search includes mutating a first neural network architecture, to form a second neural network architecture, and the mutating includes adding a residual block to the first neural network architecture or removing a residual block from the first neural network architecture.