Deep Learning Classifier Confidence Distribution Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for re-training deep learning classifiers render experimentally determined confidence score threshold values obsolete, leading to excessive and costly experimentation each time the classifier is re-trained, as they fail to constrain the outputted confidence scores during re-training, causing instability in prediction behavior.
Innovation Solution
The system uses a Gaussian mixture model based on the original training data to re-train the deep learning classifier, forcing it to maintain the original confidence score distributions, thus preserving the relevance of the previously determined threshold value.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the deep learning classifier is re-trained using existing techniques, then the classifier can learn from new data, but the confidence score threshold values become obsolete and prediction behavior becomes unstable
Solution Approach 1:
The patent implements a feedback mechanism by using the original confidence score distributions as a constraint during re-training. The loss function incorporates the difference between new and original confidence scores, creating a feedback loop that prevents the classifier from deviating too much from its original prediction behavior while still allowing it to learn from new data.
Solution Approach 2:
The patent changes the parameter space by introducing a new loss function that operates in the confidence score domain rather than just the classification accuracy domain. This parameter change allows simultaneous optimization of both adaptation to new data and preservation of original confidence distributions.
2Adaptability or versatility
If the deep learning classifier is re-trained using existing techniques, then the classifier can be updated, but excessive experimentation and resource consumption are required
Solution Approach 1:
The patent performs preliminary action by capturing and preserving the original confidence score distributions before re-training. These preserved distributions serve as a reference framework that guides the re-training process, eliminating the need for extensive post-training experimentation to determine appropriate threshold values.
Solution Approach 2:
The re-training process becomes self-service by automatically preserving the confidence score distributions through the constrained loss function. The system self-adjusts to maintain threshold value relevance without requiring external experimentation or manual calibration, thereby reducing time and resource consumption.
3Ease of manufacture
If the deep learning classifier is re-trained without constraints, then the training process is simple, but the confidence score distributions change and threshold values become invalid
Solution Approach 1:
The patent introduces asymmetry in the loss function by applying different constraints to confidence scores: the original distributions are preserved as a reference framework that cannot be changed, while the classifier is still free to adapt its predictions to new data. This asymmetric constraint structure maintains both simplicity and precision.
Data Source
AI summary
Systems and techniques that facilitate preservation of deep learning classifier confidence distributions are provided. In various embodiments, a system can access a deep learning classifier and a training dataset on which the deep learning classifier was trained. In various aspects, the system can re-train the deep learning classifier using a loss function that is based on a Gaussian mixture model constructed from the training dataset.