AI Model Retraining With Synthetic Embeddings for Misclassification Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models require significant time and resources to secure large amounts of training data for improved performance, and there is a need for more efficient methods to enhance their accuracy.
Innovation Solution
An electronic apparatus and method that utilizes a training data acquisition module to identify misclassified data, generates synthetic embedding vectors using generative adversarial networks or variational autoencoders, and updates the AI model with these vectors to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large amounts of training data are secured to improve AI model performance, then classification accuracy is improved, but time and resources required for training increase significantly
Solution Approach 1:
The patent creates synthetic copies of training data by generating artificial data samples that mimic the characteristics of real training data. This allows the AI model to be trained on expanded datasets without requiring proportional increases in time and resources, as the synthetic data can be generated efficiently through computational methods rather than manual collection and processing
Solution Approach 2:
The patent modifies parameters of existing training data to generate new synthetic training samples. By changing parameters such as noise levels, transformations, or feature variations, the system creates diverse training data that improves classification accuracy while avoiding the need to collect and process large amounts of additional real-world data
2Measurement precision
If large amounts of training data are secured to improve AI model performance, then classification accuracy is improved, but resources required for training increase significantly
Solution Approach 1:
The patent creates synthetic copies of training data that can be generated computationally without the resource-intensive processes of data collection, storage, and preprocessing associated with real-world data. This significantly reduces the energy and computational resources required while still providing sufficient training material for improving classification accuracy
Solution Approach 2:
The patent generates synthetic training data by modifying parameters of existing data samples, which requires minimal computational resources compared to collecting and processing large volumes of new real-world data. This parameter-based generation approach maintains training effectiveness while dramatically reducing resource consumption
3Loss of time
If synthetic embedding vectors are generated to correct misclassifications, then training time is reduced, but complexity of the training process increases
Solution Approach 1:
The patent implements a feedback mechanism where the AI model's classification results are analyzed to identify misclassified samples, which then become the basis for generating targeted synthetic correction data. This automated feedback loop streamlines the training process by focusing synthetic data generation only on areas needing improvement, reducing overall training time while managing complexity through systematic error correction
Solution Approach 2:
The patent enables the training system to self-correct its own deficiencies by automatically identifying misclassifications and generating appropriate synthetic training data without external intervention. This self-service approach to training optimization reduces the need for manual data curation and complex external processing, thereby reducing training time while keeping the complexity contained within the automated system
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An electronic apparatus is provided. The electronic apparatus includes a memory and a processor, wherein the processor is configured to, by executing the at least one instruction, acquire a plurality of training data; acquire a plurality of embedding vectors that are mappable to an embedding space for the plurality of training data, respectively; train an artificial intelligence model classifying the plurality of training data based on the plurality of embedding vectors, identify an embedding vector misclassified by the artificial intelligence model among the plurality of embedding vectors, identify an embedding vector closest to the misclassified embedding vector in the embedding space, acquire a synthetic embedding vector corresponding to a path connecting the misclassified embedding vector to the embedding vector closest to the misclassified embedding vector in the embedding space, and re-train the artificial intelligence model by adding the synthetic embedding vector to the training data.