Classifier Training with Knowledge Graph Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning classifiers face significant accuracy drops when transitioning to new domains due to domain changes, as seen in applications like road sign recognition models trained in one country performing poorly in another with different standards, highlighting the need for robust, context-invariant training methods.
Innovation Solution
The method employs a knowledge graph to encode prior domain knowledge, transforming it into a dense vector representation using embedding methods and guides the machine-learning system with a contrastive loss-function to adjust its data embedding space to a domain-invariant space, enabling cross-domain transfer learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classifier is trained on domain-specific data to achieve high accuracy in that domain, then accuracy in the training domain is improved, but accuracy drops when transitioning to new domains
Solution Approach 1:
The patent introduces a knowledge graph as an intermediary structure that encodes domain-invariant relationships between classes. The knowledge graph embedding layer serves as a mediator between the input data and the classifier, providing prior knowledge that guides the classifier to make domain-invariant decisions. This resolves the contradiction by decoupling the classifier from domain-specific data distributions while maintaining high accuracy through knowledge-guided predictions.
Solution Approach 2:
The patent changes the parameter representation by transforming class labels into knowledge graph embeddings (vector representations) rather than using traditional one-hot encodings. This parameter transformation allows the classifier to leverage structural relationships in the knowledge graph, enabling it to generalize across domains by relying on invariant relational patterns rather than domain-specific feature distributions.
2Measurement precision
If extensive retraining is performed to adapt a classifier to new domains, then accuracy in the new domain is improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary action by pre-constructing the knowledge graph and computing its embeddings before the classification task. This prior preparation of domain-invariant knowledge structures allows the classifier to be quickly adapted to new domains without extensive retraining, as the foundational knowledge relationships are already established and can be directly applied to new data distributions.
3Measurement precision
If a classifier is trained to be highly specialized in one domain, then accuracy in that domain is improved, but catastrophic forgetting of previous knowledge occurs
Solution Approach 1:
The patent makes the classifier universal by designing it to work across multiple domains through the knowledge graph embedding mechanism. The knowledge graph contains domain-invariant relationships that remain valid across different domains, allowing the classifier to maintain its knowledge base while adapting to new domains. This multi-functionality prevents catastrophic forgetting because the classifier relies on stable, invariant knowledge structures rather than domain-specific patterns that must be relearned.
Data Source
AI summary
A computer-implemented method for training a classifier for classifying an input signal, the input signal comprising image data, the classifier comprising an embedding part configured to determine an embedding depending on the input signal inputted into the classifier and a classification part configured to determine a classification of the input signal depending on a the embedding. The method includes: providing a first training data set of training samples, each training sample comprising an input signal and a corresponding desired classification out of a plurality of classes, providing, in a knowledge graph, additional information associated with at least one of the target classifications, providing a knowledge graph embedding method of the knowledge graph, providing a knowledge graph embedding of the knowledge graph obtained by use of a knowledge graph embedding method, training the embedding part depending on the knowledge graph embedding and the first training data set.