Distributed Edge Computer Vision Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional cloud-based computer vision systems face latency issues, making them unsuitable for real-time applications like autonomous vehicle systems and collision avoidance, due to the need for powerful computing resources and large data transmission times for model training and retraining.
Innovation Solution
Implementing distributed computer vision model training using Multi-access Edge Compute (MEC) platforms, which allow for local training and retraining of neural-network-based models at the edge, utilizing synthetic data and local synthetic training data generators to reduce latency and bandwidth requirements, enabling real-time updates and uniformity across regional MEC platforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If cloud-based computing resources are used for model training, then powerful computing capabilities are available, but latency increases and real-time performance deteriorates
Solution Approach 1:
The system segments the centralized cloud computing function into distributed edge computing nodes (MEC platforms). Each MEC platform independently performs model training and inference locally, dividing the monolithic cloud system into multiple autonomous edge units that can operate simultaneously without centralized coordination delays.
Solution Approach 2:
The system transitions from a single-dimensional centralized cloud architecture to a multi-dimensional distributed edge architecture. By adding the spatial dimension of multiple geographically distributed MEC platforms, the system achieves parallel processing capabilities that reduce latency while maintaining computing power through collective operation of edge nodes.
2Productivity
If real-time model updates are implemented, then model liveliness improves, but network bandwidth consumption increases
Solution Approach 1:
Each MEC platform maintains local model copies and performs local training with local data, eliminating the need to continuously transmit large datasets and model updates across the network. The local quality principle allows each edge node to independently update its local model without requiring synchronized global updates, dramatically reducing network bandwidth consumption while maintaining model liveliness through local adaptability.
Solution Approach 2:
The system performs preliminary model training and adaptation at the edge before deployment, rather than requiring continuous cloud-based retraining. By preparing models in advance with synthetic data and local data samples, the system reduces the frequency and volume of network transmissions needed for model updates, thereby lowering bandwidth consumption while maintaining real-time responsiveness.
3Loss of time
If distributed edge training is implemented, then latency is reduced, but system complexity increases
Solution Approach 1:
The system uses synthetic data copies and local model copies at each MEC platform to avoid complex inter-node communication and coordination. By replicating essential training components (models, synthetic data generators) at each edge node, the system achieves distributed training with minimal coordination overhead, reducing system complexity despite the distributed architecture.
Solution Approach 2:
Each MEC platform is designed to be self-sufficient, performing local model training, evaluation, and updates independently without requiring constant cloud coordination. The self-service capability of each edge node handles training latency locally while maintaining simplicity by avoiding complex centralized control mechanisms, thus reducing overall system complexity despite distributed operation.
Data Source
AI summary
Systems and methods described herein perform computer vision (CV) model training in a distributed edge network. Regional Multi-access Edge Compute (MEC) clusters are provided with a local copy of a CV model and a local synthetic training data generator. A MEC cluster receives client data requiring computer vision and applies the local copy of the CV model to the client data. The MEC cluster detects an exception to the local copy of the CV model and generates local synthetic training data for the exception. The MEC cluster updates, based on the local synthetic training data, the local copy of the CV model to form an updated local CV model. The MEC cluster sends the local synthetic training data and the updated local CV model to a central network. The central network uses the local synthetic training data to update a master CV model and any other interconnected CV models.