Joint Embedding Space Mapping for Cross-Domain Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning technologies face challenges in accurately classifying and matching data across different domains, such as video and audio, due to the need for distinct neural networks and the difficulty in representing relationships between data with different characteristics, leading to inefficiencies in training and classification accuracy.
Innovation Solution
A method involving a computing device that maps data from different domains into a common joint embedding space using a mapping neural network, generating a prediction matrix, and creating a merging dictionary to correct and refine the dataset by merging similar classes, thereby improving training efficacy and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a joint embedding space method is used to map data from different domains, then the relationship between data can be represented, but the training process requires considerable time and computational resources
Solution Approach 1:
The patent performs dataset correction before neural network training by generating a prediction matrix, identifying mislabeled data, and correcting errors in advance. This preliminary correction action reduces training time by preventing the neural network from learning incorrect patterns, directly addressing the time loss issue while maintaining the joint embedding space's ability to represent cross-domain relationships.
Solution Approach 2:
The system performs self-correction of the dataset by automatically identifying mislabeled data through the prediction matrix and correcting errors without external intervention. This self-service mechanism improves training efficiency by eliminating the need for manual dataset verification while preserving the adaptability of the joint embedding space for cross-domain data mapping.
2Measurement precision
If dataset correction is performed separately from neural network training, then data errors can be identified, but the overall training efficacy is reduced due to sequential processing
Solution Approach 1:
The patent merges dataset correction and neural network training into a unified simultaneous processing framework. The correction module and training module operate concurrently, sharing computational resources and data flows. This merging eliminates the sequential bottleneck, improving training efficacy while maintaining high accuracy in identifying and correcting data errors through the prediction matrix mechanism.
Solution Approach 2:
The system maintains continuous useful action by having the correction module continuously identify and correct errors while the training module continuously learns from the corrected data in real-time. This continuous parallel processing eliminates idle time between correction and training phases, maximizing productivity while preserving measurement precision through ongoing error detection and correction.
3Ease of manufacture
If mislabeled data is not corrected, then the dataset can be used as-is, but classification accuracy deteriorates due to training on erroneous data
Solution Approach 1:
The system automatically detects and corrects mislabeled data through the prediction matrix generated during simultaneous training, enabling self-correction without manual intervention. This maintains dataset usability by preserving the automated processing workflow while significantly improving classification accuracy by eliminating errors that would otherwise deteriorate model performance.
Solution Approach 2:
The prediction matrix provides feedback on data labeling quality by comparing predicted labels with actual labels, identifying mislabeled instances. This feedback mechanism enables continuous improvement of dataset quality during training, maintaining ease of manufacture through automated processing while enhancing reliability by correcting errors that would reduce classification accuracy.
Data Source
AI summary
An embodiment of the present disclosure discloses a method of operating a computing device for mapping data from different domains to a common joint embedding space, and the method of operating a computing device includes training a mapping neural network constituting a joint embedding space using an input dataset, generating a prediction matrix of an input dataset using the mapping neural network, and generating a merging dictionary merging classes from the prediction matrix to correct the input dataset.