Data Recognition Model Orthogonal Component Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data recognition technologies, particularly in automated systems, face challenges in accurately distinguishing between different data inputs, such as speech or images, due to limitations in generalization and orthogonality of component vectors.
Innovation Solution
A processor-implemented data recognition method that calculates an orthogonal loss and a recognition loss for temporary component vectors and embedding vectors, respectively, to train a data recognition model. This model includes a feature extraction layer, a pooling layer, and a combination layer to generate orthogonal component vectors and improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated neural network models are used for data recognition, then processing speed and automation are improved, but generalization capability and accuracy for untrained patterns deteriorate
Solution Approach 1:
The patent segments the feature space by decomposing input data into multiple orthogonal component vectors through singular value decomposition. Each component vector represents a distinct feature dimension, allowing the system to process and recognize patterns more effectively by analyzing segmented features rather than treating data as a unified whole.
Solution Approach 2:
The patent transforms the recognition approach by changing parameters from standard neural network weighting to orthogonal decomposition parameters. By using singular value decomposition to generate orthogonal component vectors and applying orthogonal loss functions, the system fundamentally alters the mathematical parameters used for pattern recognition, improving generalization capability.
2Reliability
If standard neural network training is used, then model convergence is achieved, but orthogonality of component vectors and discrimination capability deteriorate
Solution Approach 1:
The patent introduces orthogonal loss as a feedback mechanism during training. The orthogonal loss function continuously monitors and enforces orthogonality constraints on component vectors, providing feedback that guides the optimization process to maintain both convergence and discrimination capability. This feedback loop ensures that learned features remain orthogonal and discriminative throughout training.
Solution Approach 2:
The patent combines multiple loss functions (orthogonal loss and recognition loss) to create a composite objective function. This composite approach integrates the benefits of both orthogonal decomposition for discrimination and standard recognition objectives for convergence, achieving a balanced training process that maintains both reliability and adaptability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A data recognition method includes: extracting a feature map from input data based on a feature extraction layer of a data recognition model; pooling component vectors from the feature map based on a pooling layer of the data recognition model; and generating an embedding vector by recombining the component vectors based on a combination layer of the data recognition model.