Device and method for training a classifier
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning algorithms struggle to optimize non-decomposable performance metrics, which are crucial for real-world applications like optical inspection, semantic segmentation, and access control systems, due to their reliance on gradient-based methods that indirectly optimize these metrics.
Innovation Solution
A computer-implemented method for training a classifier using weighting factors that characterize non-decomposable metrics from a confusion matrix, allowing for gradient-based learning to optimize performance across a range of non-decomposable metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If gradient-based training methods are used, then training efficiency is improved, but optimization of non-decomposable metrics deteriorates
Solution Approach 1:
The patent introduces an intermediary mechanism by formulating non-decomposable metric optimization as a saddle-point problem with moment-matching constraints. This transforms the intractable direct optimization into a two-player game where gradient-based methods can operate on the relaxed problem while still converging to solutions that optimize the original non-decomposable metrics.
Solution Approach 2:
The patent changes the parameterization approach by representing conditional distributions through their moments (expectations of feature functions) rather than full probability distributions. This parameter change enables gradient-based optimization by creating a relaxed problem with tractable objective functions while preserving the ability to optimize non-decomposable metrics through the moment-matching constraints.
2Measurement precision
If full conditional distributions are optimized, then metric optimization accuracy is improved, but computational complexity deteriorates
Solution Approach 1:
The patent extracts only the essential characteristics of conditional distributions needed for metric optimization - specifically the moments (expectations of feature functions) - rather than working with full probability distributions. This extraction reduces computational complexity from exponential to polynomial while preserving the ability to optimize non-decomposable metrics accurately.
Solution Approach 2:
The patent transforms the optimization problem from the space of full conditional distributions to the space of moment vectors. This dimensionality change converts an intractable optimization over exponentially large probability spaces into a tractable optimization over polynomially-sized moment representations, enabling both accuracy and efficiency.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for training a classifier (60), particularly a binary classifier, for classifying input signals (xi) to optimize performance according to a non-decomposable metric that measures an alignment between classifications (yi) corresponding to input signals (xi) of a set of training data and corresponding predicted classifications (ŷi) of said input signals obtained from said classifier, comprising the steps of: - providing weighting factors that characterize how said non-decomposable metric depends on a plurality of terms from a confusion matrix of said classifications (yi) and said predicted classifications (ŷi); - training said classifier (60) depending on said provided weighting factors.