DNN Split Computing Control Under Unstable Wireless Links
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing communication systems face challenges in efficiently distributing deep neural network computations across communication terminals, cloud servers, and communication networks, leading to increased computation and communication delays due to unstable network quality and varying device capabilities.
Innovation Solution
An information processing device dynamically adjusts computation handling across multiple devices by determining whether to transmit intermediate results or execute computations locally, optimizing computation distribution based on device capabilities and network conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If computation of DNN is distributed to both communication terminal and cloud server, then computation time is reduced, but communication volume increases and delay occurs due to unstable wireless communication quality
Solution Approach 1:
The patent dynamically adjusts the computation distribution between communication terminal and cloud server based on real-time network conditions. The system monitors communication quality and adapts the splitting point of DNN computation accordingly, transitioning between edge computing and cloud computing modes to maintain optimal performance under varying network reliability
Solution Approach 2:
The system changes the parameter of computation splitting point in the DNN architecture based on network conditions. By adjusting which layer of the neural network is executed locally versus remotely, the system optimizes the balance between computation time and communication reliability according to current wireless communication quality
2Power
If computation of DNN is concentrated on cloud server, then computation capability is sufficient, but communication volume increases and delay amount may exceed allowable limit
Solution Approach 1:
The patent segments the DNN computation into multiple layers and distributes them between the communication terminal and cloud server. The communication terminal executes early layers locally while later layers are processed by the cloud server, reducing the amount of data that needs to be transmitted and thereby reducing communication delay while maintaining sufficient computation capability
3Quantity of substance
If computation of DNN is concentrated on communication terminal, then communication volume is reduced, but computation time and power consumption increase
Solution Approach 1:
The DNN computation is segmented into multiple layers with the terminal handling early layers and the cloud server handling later layers. This segmentation reduces communication volume compared to full cloud processing while avoiding the excessive computation time and power consumption of complete local processing
4Productivity
If computation of DNN is distributed to multiple devices, then computation load is balanced, but computation may take longer than assumed due to unstable load and communication quality
Solution Approach 1:
The system implements feedback mechanisms to monitor the actual computation time and communication quality in real-time. Based on this feedback, the computation splitting point and distribution strategy are dynamically adjusted to ensure that the total computation time remains within expected limits despite variations in device load and network conditions
Data Source
AI summary
To an information processing device or the like that reduces the time required to transmit a computation result as a response while causing a computation on the basis of a DNN to be distributed. According to an aspect of the present disclosure, there is provided an information processing device that handles a part of a series of computations in a deep neural network. The information processing device determines whether or not to transmit, to a first communication device, a result of an intermediate computation in the series of computations in the deep neural network, in a case where transmission to the first communication device is not determined, transmits a result of at least a part of a computation included in a first range among the series of computations to a second communication device that is a computation handler of a second range following the first range, and, in a case where transmission to the first communication device is determined, executes a computation for transmission to the first communication device by using a result of the intermediate computation in the series of computations in the deep neural network, and transmits a result of executing the computation to the first communication device.