Model Output Device Greedy Optimization for Classifier Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for extracting representative examples from complex classifiers, such as those used in deep learning, require high calculation costs due to the combinatorial optimization problem, which can take several days to solve even with parallel calculators like GPUs, making them inefficient for large-scale models.
Innovation Solution
The model output device employs an elimination greedy method that selects the second-most-similar interpretable model for each test data point and excludes the model with the smallest classification error, reducing the calculation cost and improving extraction accuracy by considering both similarity and error viewpoints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If combinatorial optimization is used to extract representative examples from complex classifiers, then extraction accuracy is improved, but calculation cost increases significantly
Solution Approach 1:
The patent segments the combinatorial optimization problem into two separate optimization processes: one for selecting test data points and another for selecting approximation models. This division allows each sub-problem to be solved more efficiently using greedy optimization with local search, avoiding the exponential complexity of solving the full combinatorial problem at once while maintaining extraction accuracy.
2Measurement precision
If combinatorial optimization is used to extract representative examples, then extraction accuracy is improved, but processing time increases to several days
Solution Approach 1:
The patent divides the time-consuming combinatorial optimization into two sequential greedy optimization steps with local search refinement. This segmentation reduces processing time from several days to a practical duration while preserving extraction accuracy by addressing data selection and model selection as separate, manageable optimization tasks.
Solution Approach 2:
The patent performs preliminary selection of test data points and approximation models using greedy optimization before applying local search refinement. This preliminary action filters the search space in advance, making the subsequent optimization more efficient and reducing overall processing time while maintaining accuracy.
3Reliability
If all approximation models are retained for analysis, then completeness of analysis is improved, but device complexity increases
Solution Approach 1:
The patent extracts and selects only the most relevant approximation models from the full set of generated models by optimizing the selection process. This extraction of essential models reduces analysis complexity while maintaining completeness by ensuring that the selected subset captures the necessary information through targeted optimization rather than exhaustive analysis of all models.
Data Source
AI summary
A storage medium storing a program that causes a computer to execute a process, the process includes obtaining a plurality of test data used for analysis of a classifier; obtaining a plurality of approximation models approximated to the classifier, each of separation boundaries of the plurality of approximation models being approximated to a different part of a separation boundary of the classifier; specifying, for each of the approximation models, a test data having a highest similarity from among the test data; specifying, for each specified test data, an approximation model having a second highest similarity with the specified test data from among the approximation models; selecting, from among the approximation models, an exclusion target model whose error with the specified approximation model is the smallest for each specified test data; and outputting the plurality of approximation models from which the exclusion target model is excluded for each specified test data.