Tree-Form AI Model Architecture for Edge Memory and Power Use
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in efficiently managing multiple AI models on resource-constrained edge devices, leading to exorbitant memory usage, power consumption, and latency, particularly in applications like camera functions, due to the need for numerous AI models to perform diverse tasks without effective scaling or integration.
Innovation Solution
A tree-form AI model architecture is developed, comprising a trunk model for common operations and branch models for specific tasks, optimized through Neural Architecture Search (NAS) to reduce floating-point operations and memory usage while maintaining accuracy, allowing for scalable addition of new tasks via transfer learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple independent AI models are deployed for different tasks, then task-specific accuracy is improved, but memory usage and power consumption increase significantly
Solution Approach 1:
The patent merges multiple independent AI models into a single unified model where different tasks are represented as separate branches sharing a common trunk. This consolidation reduces the total memory footprint while maintaining task-specific accuracy through specialized branch modules.
Solution Approach 2:
The unified AI model achieves multi-functionality by incorporating multiple task-specific branches that can be selectively activated. The shared trunk provides common feature extraction capabilities that serve all tasks, reducing redundant memory usage while maintaining specialized performance for each task.
2Adaptability or versatility
If multiple independent AI models are deployed for different tasks, then task coverage is improved, but power consumption increases
Solution Approach 1:
By merging multiple task-specific models into one unified model with shared components, the system reduces the total computational overhead and power consumption while maintaining comprehensive task coverage through the multi-branch architecture.
Solution Approach 2:
The system dynamically activates only the necessary branches for the current task, avoiding the continuous power consumption that would result from running all independent models regardless of which task is currently needed.
3Measurement precision
If multiple independent AI models are deployed for different tasks, then task specialization is improved, but model switching time increases
Solution Approach 1:
Merging multiple specialized models into a single unified model with task-specific branches eliminates the need to load and switch between separate model files. The system can rapidly switch between tasks by activating different branches within the same model structure, reducing switching time while maintaining task specialization.
4Quantity of substance
If a single unified AI model is created for multiple tasks, then memory usage is reduced, but task-specific performance may deteriorate
Solution Approach 1:
The unified model is segmented into a shared trunk for common operations and separate branches for task-specific operations. This segmentation allows the model to maintain task-specific performance by dedicating separate computational paths for each task while sharing common feature extraction resources in the trunk.
Solution Approach 2:
Different parts of the unified model have different qualities optimized for their specific functions. The trunk provides general-purpose feature extraction, while each branch is specialized for its specific task, ensuring high performance for each task despite the unified structure.
Data Source
AI summary
A method, performed by a device, includes identifying data for multiple tasks performed based on different Artificial Intelligence (AI) models; configuring a single tree-form AI model comprising a trunk model and multiple branch models, where each branch model performs a different task; and training the model using datasets for the various tasks. The trunk model performs common operations and is heavier than the branch models. The model architecture and task weightages are determined using Neural Architecture Search to optimize resource usage by decreasing floating-point operations and memory usage in branch models while increasing them in the trunk model to improve overall accuracy. The method supports adding new branch models for new tasks using transfer learning without altering the trunk model. A complementary method involves loading the trunk model into memory, identifying a target task, and loading only the corresponding branch model to efficiently perform the task.