Digital Asset Tagging via Pre-trained Model Adaptation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional digital asset management systems face challenges in accurately propagating generic tags to custom tags across different vocabulary sets, leading to reduced accuracy and efficiency in digital asset searches, especially when dealing with diverse digital assets like images of cars with varying visual classes.

Innovation Solution

The system employs machine learning models trained on a generic vocabulary set to generate custom tags by extracting last layer activations and using them to train sub-models for specific visual classes, allowing for efficient tagging across different vocabulary sets without the need for retraining from scratch, even with limited training data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional service provider systems re-train base neural network models from scratch to propagate generic tags to custom tags, then tagging accuracy for custom vocabulary sets is improved, but computational cost and time consumption increase significantly

Engineering Contradiction:

Improvetagging accuracyVSAvoidtraining time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system pre-trains base neural network models on large datasets using a generic vocabulary set before deployment. These pre-trained models serve as foundational components that can be rapidly adapted to custom vocabulary sets without requiring retraining from scratch, thus reducing training time while maintaining tagging accuracy

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system changes the parameter configuration of existing models by adjusting vocabulary mappings and tag propagation parameters rather than retraining the entire model. This allows the same base model to adapt to different custom vocabulary sets efficiently, reducing both computational cost and training time

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

Engineering Contradiction:

Improvetagging accuracyVSAvoidcomputational resources

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

Base neural network models are pre-trained on large datasets using a generic vocabulary set before deployment. These pre-trained models serve as foundational components that can be rapidly adapted to custom vocabulary sets without requiring retraining from scratch, thus reducing computational resources while maintaining tagging accuracy

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system creates and uses copies of pre-trained base models for different custom vocabulary sets rather than retraining original models each time. This copying approach significantly reduces computational resource consumption while maintaining the accuracy benefits of trained models

Inventive Principle:

Principle #26Copying

3Adaptability or versatility

If brute-force techniques using similarity metrics are used to propagate tags across digital assets, then coverage of tagging is improved, but computational feasibility deteriorates for large numbers of digital assets

Engineering Contradiction:

Improvetag propagation coverageVSAvoidcomputational feasibility

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The system segments the tag propagation process into two stages: first, base models generate generic tags for digital assets; second, trained propagation models selectively map generic tags to custom tags based on learned relationships. This segmentation avoids the computational infeasibility of brute-force comparison while maintaining comprehensive tag coverage

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system introduces generic tags as an intermediary layer between the base model and custom vocabulary sets. This intermediary approach enables efficient tag propagation by leveraging the base model's generic tag generation capability and the propagation model's learned mapping relationships, avoiding direct brute-force comparison

Inventive Principle:

Principle #24Intermediary (Mediator)

4Measurement precision

If base neural network models are re-trained from scratch to accommodate custom vocabulary sets, then tagging accuracy for specific visual classes is improved, but the number of training data required increases significantly

Engineering Contradiction:

Improvetagging accuracyVSAvoidtraining data volume

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The system changes the parameter configuration of existing models by adjusting vocabulary mappings and tag propagation parameters rather than retraining the entire model. This allows the same base model to adapt to different custom vocabulary sets efficiently with minimal training data, maintaining tagging accuracy

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11301506B2Automated digital asset tagging using multiple vocabulary sets

Publication Date: 2022.04.12 ADOBE INC

AI summary

Automated digital asset tagging techniques and systems are described that support use of multiple vocabulary sets. In one example, a plurality of digital assets are obtained having first-vocabulary tags taken from a first-vocabulary set. Second-vocabulary tags taken from a second-vocabulary set are assigned to the plurality of digital assets through machine learning. A determination is made that at least one first-vocabulary tag includes a plurality of visual classes based on the assignment of at least one second-vocabulary tag. Digital assets are collected from the plurality of digital assets that correspond to one visual class of the plurality of visual classes. The model is generated using machine learning based on the collected digital assets.