Train-once-for-all Personalization Framework for Edge AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, especially large-scale generic models, are inefficient for personalization and resource-constrained devices due to high computational and memory requirements, and lack scalability for user-specific tasks, leading to suboptimal performance in classifying specific objects.
Innovation Solution
A 'train-once-for-all' personalization framework that learns a set of basis models and a mixer predictor, allowing for real-time combination of model weights to create personalized models based on task descriptions, enabling efficient adaptation and deployment on resource-limited devices without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large-scale generic machine learning model is used to handle multiple tasks, then the model can classify thousands of objects, but the model size and computational resources become too large for deployment on resource-constrained client devices
Solution Approach 1:
The generic model is segmented into multiple task-specific sub-models, each trained to handle specific user tasks. Instead of deploying one large generic model, the system divides it into smaller specialized components that can be selectively activated based on user needs, reducing the model size deployed on client devices while maintaining versatility.
Solution Approach 2:
The patent extracts only the necessary task-specific components from the large generic model based on individual user requirements. Rather than deploying the entire generic model, the system identifies and extracts relevant sub-models for each user's specific tasks, reducing the deployed model size while preserving the needed functionality.
2Measurement precision
If specialized models are trained for each user's specific requirements, then personalized accuracy is improved, but the computation for training grows linearly with the number of users and their requests
Solution Approach 1:
The system performs preliminary training by pre-training a generic model on diverse data and then pre-segmenting it into task-specific sub-models before actual user deployment. This preliminary preparation work is done once, and then user-specific personalization is achieved by selectively combining these pre-prepared components, avoiding the need to retrain from scratch for each user and thus improving scalability.
Solution Approach 2:
The patent uses copying by creating task-specific sub-models as copies or variants of the generic model structure. These sub-models are trained on specific task data and can be selectively activated for different users. This copying approach allows the system to serve multiple users with different task requirements without training entirely new models for each user, improving training scalability while maintaining personalized accuracy.
3Adaptability or versatility
If a generic model is updated to cover newer tasks or changed datasets, then the model remains versatile, but the update cost from processing resources and time becomes prohibitively expensive
Solution Approach 1:
The model is segmented into independent task-specific sub-models that can be updated separately. When new tasks or datasets are introduced, only the relevant sub-models need to be updated rather than retraining the entire generic model. This segmentation enables incremental updates, reducing both processing resources and time required for model adaptation.
Solution Approach 2:
The system implements dynamic model composition where the set of active sub-models can be flexibly adjusted based on current user needs and new tasks. When new tasks emerge, the system can dynamically add, remove, or update specific sub-models without affecting the entire model system, enabling rapid adaptation with minimal update time and resources.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The technology relates to providing personalized neural network-based models according to user input, which can be generated upon request or otherwise as needed. This may include receiving, by one or more processors of a computing device, input corresponding to a task description. Then the input corresponding to the task description is encoded into a set of text embeddings. Based on this, the system applies mixer prediction to the set of text embeddings to generate a set of mixers and learns a set of basis models according to the set of mixers. The set of basis models are combined to form a single personalized model corresponding to the task description. This personalized model can then be used in video understanding, quality assessment, providing a recommendation, performing a classification, or performing a search.