Second Classifier Reduces Errors in Trained ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in improving performance due to high computational complexity and classification errors, particularly in reducing the number of errors in low complexity classifiers.
Innovation Solution
The solution involves adding a second classifier with a non-differentiable objective function to directly minimize classification errors and utilizing soft probabilities from a higher complexity model to improve the performance of a trained machine learning model, without retraining the original model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a low complexity classifier is used, then computational complexity is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent combines a low complexity classifier with a error correction module that processes classification errors. The error correction module receives incorrect classifications from the low complexity classifier and applies additional processing to correct them, effectively merging two functional components to achieve both low complexity and high accuracy.
Solution Approach 2:
The error correction module acts as an intermediary between the low complexity classifier and the final classification output. It receives the classifier's predictions, identifies errors through its objective function, and corrects them before producing the final result, thereby improving accuracy without increasing the complexity of the main classifier.
2Reliability
If a higher complexity model is used to improve accuracy, then classification errors are reduced, but training time and computational resources increase
Solution Approach 1:
The patent extracts only the essential error correction functionality from a complex model and implements it as a separate, lightweight module. Instead of using a full higher complexity model, it extracts the error correction capability and implements it with minimal parameters and computational requirements, achieving accuracy improvement without proportional increase in training time.
Solution Approach 2:
The error correction module uses a simplified objective function that is computationally inexpensive to evaluate and optimize. This allows for rapid training and iteration, making the accuracy improvement achievable with minimal time investment compared to training a complete high-complexity model from scratch.
3Reliability
If classification errors are directly minimized, then accuracy is improved, but model complexity increases
Solution Approach 1:
The patent segments the classification system into two distinct parts: a low complexity classifier for initial predictions and a separate error correction module for improving accuracy. This segmentation allows each component to be optimized independently, with the error correction module focusing solely on minimizing classification errors without increasing the complexity of the main classifier.
Solution Approach 2:
The error correction module applies localized improvements only where needed - specifically for correcting classification errors. Rather than increasing the complexity of the entire model, it applies additional processing only to incorrect classifications, maintaining low overall complexity while achieving high accuracy where it matters most.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A method for improving performance of a trained machine learning model includes adding a second classifier with a second objective function to a first classifier with a first objective function. Rather than minimizing a function of errors for the first classifier, the second objective function is used to directly reduce the number errors of the first classifier.