Neural Network Classifier Zero-Shot Classification via Knowledge Graph
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classifiers trained for measurement data classification struggle with zero-shot classification, where they need to identify classes they have not been trained on, without requiring additional training samples or altering their behavior.
Innovation Solution
A method using a neural network classifier with a feature extraction section and a knowledge graph to determine activated neurons and their corresponding attributes, allowing the classifier to predict classes it has not seen during training by linking new classes to attributes and neurons, enabling zero-shot classification without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a classifier is trained only on a finite set of classes with ground truth labels, then the classifier achieves reliable classification for those known classes, but it cannot classify measurement data belonging to unseen classes (zero-shot classification capability is lost)
Solution Approach 1:
The patent introduces an intermediary knowledge graph that connects classes and attributes. Instead of directly mapping measurement data to classes, the system uses activated neurons to determine present attributes, then uses the knowledge graph to infer classes from those attributes. This intermediary structure enables zero-shot classification while maintaining reliability for known classes.
Solution Approach 2:
The patent adds a new dimension to the classification process by introducing attribute space as an intermediate representation. Rather than operating solely in class space, the system transforms the problem into attribute space through neuron activation patterns, then maps back to class space using the knowledge graph. This dimensional transformation enables handling of unseen classes.
2Adaptability or versatility
If additional training samples are used to improve zero-shot classification performance, then the classifier can recognize more classes, but the training time and computational resources increase
Solution Approach 1:
The patent performs preliminary action by pre-construction of the knowledge graph linking classes and attributes before classification tasks. This pre-established semantic structure allows the classifier to handle unseen classes without requiring additional training samples, as the reasoning capability is prepared in advance through the knowledge graph rather than through extensive training.
Solution Approach 2:
The system uses its own internal neuron activation patterns to automatically determine which attributes are present in the measurement data. This self-determined attribute information is then used with the knowledge graph to infer classes, eliminating the need for external training samples or manual annotation for new classes.
3Adaptability or versatility
If the classifier architecture is modified to enable zero-shot classification, then the classifier can handle unseen classes, but the device complexity increases
Solution Approach 1:
The knowledge graph serves as a universal structure that can accommodate any number of classes and attributes without requiring changes to the classifier architecture. The same feature extraction section and knowledge graph mechanism handle both seen and unseen classes uniformly, providing multi-functionality without increasing structural complexity.
4Adaptability or versatility
If retraining the classifier is performed to include new classes, then the classifier accuracy improves for those classes, but certification in high-risk applications like autonomous driving is compromised
Solution Approach 1:
The knowledge graph acts as a certified intermediary that bridges the classifier and new classes. Since the knowledge graph can be independently verified and the classification logic is transparent (neuron activation → attribute determination → class inference), the system can be certified without retraining, as the core classifier remains unchanged while the reasoning layer handles new classes.
Data Source
AI summary
A method for operating at least one trained classifier for measurement data. The classifier comprises a neural network with at least one feature extraction section and at least one classification section. The method includes: processing a record of measurement data with at least the feature extraction section of the classifier; determining a set of neurons in the feature extraction section that are activated by said processing; determining, from a given correspondence between activated neurons and attributes, a set of attributes whose presence in a scene captured by the measurement data is indicated by the activated neurons; comparing attributes to which classes are linked by a given knowledge graph with said determined set of attributes; and evaluating, from the result of this comparison, at least one estimated class as a class to which the scene captured by the record of measurement data is likely to belong.