Genetic Programming for AI Accelerator Neural Network Architecture Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing evolution algorithms for AI accelerators are time-consuming and computationally expensive, making them difficult to adapt to the evolving needs of AI accelerators and requiring further optimization.
Innovation Solution
The proposed optimizing method for AI accelerators uses genetic programming to search for target neural network architectures, employing a tree structure, elite selection, and acquired inheritance to reduce computational costs and improve efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If evolutionary algorithms are used to search for optimal neural network architectures, then the global optimal solution can be found, but the execution time and computational cost increase significantly
Solution Approach 1:
The patent applies preliminary action by performing data preprocessing, feature extraction, and search space definition before the evolutionary search begins. The search space is pre-defined with constraints and boundaries, and the fitness evaluation framework is prepared in advance, allowing the evolutionary algorithm to focus only on optimizing within predefined parameters rather than exploring the entire solution space from scratch.
Solution Approach 2:
The patent segments the neural network architecture search into multiple independent components: search space definition, fitness function design, population initialization, and evolutionary operations. Each component can be independently optimized and configured. The search process is divided into generations with intermediate evaluations, allowing parallel computation and reducing overall execution time.
2Measurement precision
If manual debugging is used to design neural network models, then researcher expertise can be applied, but the process consumes excessive time especially as model size increases
Solution Approach 1:
The patent implements self-service by enabling the system to automatically perform neural network architecture search and optimization without requiring manual debugging. The evolutionary algorithm autonomously explores the search space, evaluates fitness, and converges to optimal architectures. The system self-adjusts parameters and configurations based on performance feedback, eliminating the need for researcher intervention in the debugging process.
Solution Approach 2:
The patent applies parameter changes by systematically varying neural network architecture parameters (layer sizes, activation functions, connectivity patterns) through the evolutionary process. Instead of manual parameter tuning, the system automatically modifies parameters across generations, exploring combinations that optimize performance while reducing the time researchers would spend on manual debugging.
3Reliability
If the search space for weight parameters and feature parameters is expanded to improve model performance, then better neural networks can be achieved, but the time required to debug parameters increases exponentially
Solution Approach 1:
The patent applies preliminary action by pre-defining the search space boundaries, valid parameter ranges, and architectural constraints before the evolutionary search begins. This preprocessing step structures the parameter space in a way that guides the search toward promising regions while avoiding invalid configurations, reducing the effective search space without sacrificing performance potential.
Solution Approach 2:
The patent replaces the mechanical manual debugging process with an automated evolutionary computation system. Instead of researchers manually adjusting parameters one by one, the system uses algorithmic operations (selection, crossover, mutation) to automatically explore and optimize parameters, substituting human cognitive effort with computational processes that can handle large parameter spaces efficiently.
Data Source
AI summary
The present invention discloses an optimizing method for an AI accelerator and an AI accelerator. The optimizing method optimizes the AI accelerator through obtaining target neural network architecture by genetic programming, and includes: preparing required raw data, removing abnormal data, annotating based on different data types to obtain annotated data, and selecting part of the annotated data as a training set; determining search space of genetic programming, defining function set and terminal set of genetic programming, and performing preprocessing, extracting features, concatenating features, regressing and result outputting on the annotated data; defining fitness function used in genetic programming to search for optimal individuals; and the training set performing to search for obtaining the target neural network architecture. The present invention addresses the issues of interpretability and understandability in traditional neural network generation by leveraging the encoding capabilities of genetic programming.