Scalable Computer Vision Offloading With Dynamic DNN Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mobile devices face challenges in hosting hardware and energy storage necessary for timely execution of computationally and memory-intensive Deep Neural Networks (DNNs) for computer vision tasks, leading to inefficiencies in resource utilization and task execution.
Innovation Solution
A system is proposed that includes a computing resource pool, a DNN repository, and a controller that determines the optimal DNN configuration by applying tasks to a graph model representing solution paths through a DNN structure. This system enables dynamic DNN construction by selecting and activating appropriate DNN blocks from the repository, based on the required computing resources and task admission ratios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If mobile devices host hardware and energy storage for DNN execution, then task execution speed improves, but device complexity and energy consumption increase beyond acceptable limits
Solution Approach 1:
The patent extracts the DNN execution functionality from mobile devices and relocates it to edge servers. The mobile device offloads CV tasks to edge servers that host the DNN models and computing resources, allowing fast execution without increasing device complexity. This is achieved through task offloading where the computational workload is separated from the end-user device.
Solution Approach 2:
The patent introduces edge servers as intermediary components between mobile devices and cloud infrastructure. These edge servers act as mediators that provide DNN execution capabilities to mobile devices without requiring the devices themselves to host complex hardware. The edge server receives offloaded tasks, executes them using hosted DNN models, and returns results to the mobile device.
2Power
If mobile devices host hardware for DNN execution, then processing capability improves, but energy storage requirements exceed device capacity
Solution Approach 1:
The patent extracts the energy-intensive DNN processing function from mobile devices and relocates it to edge servers with adequate power supply. This allows high processing capability to be achieved without burdening mobile device energy storage systems, as the computational workload is performed remotely at the edge infrastructure.
3Device complexity
If conventional task offloading is used, then mobile device resource requirements decrease, but edge server resource utilization becomes inefficient
Solution Approach 1:
The patent implements dynamic DNN configuration at edge servers that adapts to varying task requirements and resource availability. The system dynamically selects and activates appropriate DNN blocks from a repository based on task characteristics and current edge server load, optimizing resource utilization. This dynamic approach allows the edge server to efficiently handle diverse CV tasks while maintaining low resource requirements at mobile devices.
4Measurement precision
If fixed DNN models are deployed at edge, then task processing accuracy is maintained, but system adaptability to diverse tasks decreases
Solution Approach 1:
The patent segments DNN models into modular blocks that can be independently selected and activated. Instead of deploying fixed complete models, the system maintains a repository of DNN blocks and dynamically assembles them into complete models based on task requirements. This segmentation enables the system to maintain accuracy for specific tasks while adapting to diverse task types by selecting appropriate block combinations.
Solution Approach 2:
The patent implements dynamic DNN configuration that adapts to diverse CV tasks by selectively activating appropriate DNN blocks from the repository. The system dynamically determines which blocks to activate based on task characteristics, enabling versatility across different task types while maintaining the accuracy benefits of specialized DNN configurations for each task category.
Data Source
AI summary
Tasks are distributed among a computing resource pool for operating a deep neural network (DNN) repository comprising a plurality of DNN models. A DNN configuration to process the task is determined based on an identification of computing resources required to process the task. A subset of the computing resource pool is allocated to execute the task based on the DNN configuration. A selected set of DNN blocks from the DNN repository is activated based on the DNN configuration, and a device transmit input data to the subset of the computing resource pool to execute the task via the selected set of DNN blocks.