METHOD AND APPARATUS WITH AI MODEL TRAINING USING DOMAIN SIMILARITY

Smart overview of the Invention

The patent describes a method and apparatus for enhancing the training of a pre-trained AI model by leveraging domain similarity. The process involves generating two images from different domains, calculating gradients of associated loss functions, and determining their similarity. This similarity is used to update the AI model, optimizing its performance across diverse domains.

Technical Approach

The method calculates a first gradient for an image from the pre-training domain and a second gradient for an image from the target domain. By assessing the similarity between these gradients, the model's parameters are updated. This approach allows for efficient adaptation of the model to new domains, potentially reducing computational resources and training time compared to traditional methods.

Apparatus Configuration

The apparatus comprises processors and memory where instructions are stored to execute the training method. It generates image pairs from pre-training and target domains and evaluates the similarity between these domains. The model with the highest similarity is selected for further training, facilitating effective transfer learning.

Gradient Adjustment

Updating the model involves adjusting the gradient of each layer based on the similarity between the first and second gradients. This adjustment is calculated using a product of the similarity and the second gradient, ensuring that the model's learning is aligned with the target domain's characteristics.

System Implementation

The system is designed to adapt a pre-trained AI model to a target domain by generating images and calculating correlation values. These values guide the tuning process, updating the model's parameters to improve its performance in the new domain. The approach ensures that the model retains its pre-trained knowledge while effectively learning new domain-specific features.