Automated Hyperparameter Optimization for Computer Vision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning systems for computer vision require manual hyperparameter adjustment, leading to sub-optimal performance and limited applicability across different data sets, especially those with varying numbers of classes or domains.
Innovation Solution
An automated method using BOHB for hyperparameter optimization, combined with a decision tree and meta-features, to determine optimal parameterizations for machine learning systems, enabling robust and domain-independent learning across diverse data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual hyperparameter adjustment is used, then expert experience can guide the process, but the system converges to sub-optimal solutions and requires significant power
Solution Approach 1:
The system performs self-service by automatically determining optimal hyperparameters through the automated hyperparameter determination unit, eliminating the need for manual expert adjustment while achieving optimal solutions that reduce power consumption in the machine learning system
Solution Approach 2:
The system changes parameters by systematically varying hyperparameters within defined value ranges and using optimization algorithms to identify the optimal parameter configuration that maximizes performance while minimizing energy consumption
2Reliability
If hyperparameter optimization is performed for each training data set, then optimal results can be achieved, but the process requires extensive time and computational resources
Solution Approach 1:
The system performs preliminary action by pre-determining optimal hyperparameters through the automated hyperparameter determination unit before actual training begins, storing these optimized parameters for reuse across similar data sets to avoid repeated optimization processes
Solution Approach 2:
The system achieves universality by creating a reusable hyperparameter configuration that can be applied across multiple different training data sets and machine learning algorithms, allowing one optimization result to serve multiple purposes and reduce overall optimization time
3Ease of operation
If manual hyperparameter adjustment is performed, then some control over the process is maintained, but optimal parameterization cannot be achieved and power losses occur
Solution Approach 1:
The system performs self-service by automatically determining optimal hyperparameters through the automated hyperparameter determination unit, eliminating the need for manual expert adjustment while achieving optimal solutions that reduce power consumption in the machine learning system
Solution Approach 2:
The system replaces manual mechanical adjustment with an automated computational system that uses optimization algorithms and automated determination units to select hyperparameters, substituting human effort with an efficient automated process that achieves better results with less energy
4Adaptability or versatility
If hyperparameter optimization is performed without automation, then the process can be controlled manually, but the system is not applicable across different data sets and domains
Solution Approach 1:
The system achieves universality by creating a reusable hyperparameter configuration that can be applied across multiple different training data sets and machine learning algorithms, allowing one optimization result to serve multiple purposes and reduce overall optimization time
Solution Approach 2:
The system performs self-service by automatically determining optimal hyperparameters through the automated hyperparameter determination unit, eliminating the need for manual expert adjustment while achieving optimal solutions that reduce power consumption in the machine learning system
Data Source
AI summary
Computer-implemented method for creating a system, which is suitable for creating in an automated manner a machine learning system for computer vision. The method includes: providing predefined hyperparameters; determining an optimal parameterization of the hyperparameters using BOHB (Bayesian optimization (BO) and Hyperband (HB)) for a plurality of different training data sets; assessing all optimal parameterizations on all training data sets of the plurality of different training data sets with the aid of a normalized metric; creating a matrix, the matrix including the evaluated normalized metric for each parameterization and for each training data set; determining meta-features for each of the training data sets; optimizing a decision tree, which outputs as a function of the meta-features and of the matrix which of the optimal parameterization using BOHB is a suitable parameterization for the given meta-features.