Neural Model Quality Assurance With Generative Prototypes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing discriminative neural networks struggle to reliably classify input data records that deviate significantly from the training data, leading to issues such as overfitting and inaccurate classifications, particularly in medical imaging where technical parameters like contrast range and image resolution vary.
Innovation Solution
A system combining a discriminative neural network classifier with a generative neural network model-based sample generator, trained on the same data, generates artificial prototypes to determine the reliability of classifications by assessing the similarity of input data records to these prototypes, using loss functions or similarity metrics to establish a parameter space for reliable classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a discriminative neural network is trained on training data records, then it can classify objects within the training data distribution, but it fails to reliably classify input data records that deviate significantly from the training data
Solution Approach 1:
The patent applies preliminary action by generating artificial prototype data records before actual classification tasks. These prototypes are created using a generative neural network trained on the training data distribution, establishing a reference framework in advance that helps evaluate whether new input data falls within the reliable classification range, thus preparing the system beforehand for out-of-distribution detection
Solution Approach 2:
The patent introduces an intermediary mechanism by using artificial prototype data records as mediators between the training data and new input data. These prototypes serve as reference points that bridge the gap, allowing the system to compare new inputs against established prototypes to determine if they deviate significantly, thus enabling reliable classification boundaries without direct exposure to out-of-distribution data
2Measurement precision
If the classifier is trained to recognize patterns in training data, then it achieves good accuracy on training data, but it overfits and produces inaccurate classifications on novel data with different technical parameters
Solution Approach 1:
The patent applies copying by creating artificial copies of the training data distribution through generative neural networks. These synthetic prototype data records replicate the statistical properties and technical parameter ranges of the training data, providing a safe reference framework that prevents overfitting to specific training samples while maintaining accurate pattern recognition for in-distribution data
3Adaptability or versatility
If technical parameters like contrast range and image resolution vary in input data, then the classifier encounters out-of-distribution data, but existing systems lack a mechanism to identify such variations and maintain reliable classification
Solution Approach 1:
The patent applies parameter changes by systematically varying technical parameters (such as contrast range, image resolution, and other acquisition parameters) when generating artificial prototype data records. This creates a comprehensive reference framework that encompasses the expected range of parameter variations, enabling the system to identify when new inputs fall outside this range and maintain reliable classification by detecting these parameter deviations
Data Source
AI summary
The invention relates to a system which, on the one hand, has a classifier that is formed by a discriminative neural network and that implements a binary class model or a multi-class model. On the other hand, the system has a model-based sample generator that is formed by a generative neural network. Both the classifier and the model-based sample generator are trained—for a corresponding class—with the same training data records and therefore embody models that correspond to one another for this class.The invention also relates to a method for determining a quality criterion for input data records for a classifier with a discriminative neural network. The classifier has been trained with training data records and represents a classification model for a class.According to the method, a model-based sample generator with a generative neural network is initially provided and trained with the same training data records that were used to train the classifier.Subsequently, by means of the trained model-based sample generator and an input data record based on random values, an artificial data record is generated, which is representative of the classification model embodied by the classifier.The artificial data record generated by the trained generator, or at least a parameter derived from it, is used to test the input data records as to their suitability for classification or regression.