Neural Network Update Encoding for Layer-Level Data Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems transmitting neural network (NN) structure updates transmit unnecessary information, leading to an increased data size, as both updated and non-updated layer information are sent, which hampers efficient data compression and processing.
Innovation Solution
A data processing device with encoding circuitry that generates encoded data by identifying and encoding only the updated layer information, along with model and layer header information, and layer structure details, using flags to indicate updates, thereby reducing the data size transmitted.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If information about both updated and non-updated layers is transmitted, then clients can receive complete NN structure information, but the data size to be transmitted increases
Solution Approach 1:
The patent extracts only the necessary information for NN structure updates by identifying and transmitting solely the updated layer information, excluding non-updated layer data. This selective extraction reduces the transmitted data size while maintaining the completeness of update information, directly resolving the contradiction between data completeness and data size.
Solution Approach 2:
The patent segments the NN structure information into individual layer units, allowing the transmission system to selectively send only the updated layer segments rather than the complete NN structure. This segmentation enables efficient partial updates, reducing overall data transmission volume while ensuring all necessary updates are delivered.
2Reliability
If complete NN structure information is transmitted, then clients can maintain accurate models, but the processing load on the transmission system increases
Solution Approach 1:
The system extracts and transmits only the essential update information regarding changed layers, removing unnecessary redundant data about unchanged layers. This extraction approach maintains model accuracy on the client side while significantly reducing the processing load on the transmission system, thereby improving data transmission efficiency.
3Reliability
If all layer information is encoded and transmitted, then clients can ensure model consistency, but data compression efficiency decreases
Solution Approach 1:
The patent applies extraction by identifying and encoding only the updated layer information, removing redundant encoding of unchanged layers. This selective encoding process maintains model consistency on the client side through precise update information while achieving superior data compression ratios by eliminating unnecessary data.
Solution Approach 2:
The system performs partial encoding action by encoding only the necessary updated portions of the NN structure rather than the complete structure. This partial action approach ensures model consistency through complete update information while maximizing compression efficiency by avoiding excessive encoding of unchanged data.
Data Source
AI summary
There are included a data processing unit that trains a neural network; and an encoding unit that generates encoded data in which model header information for identifying a model of the neural network, layer header information for identifying one or more layers of the neural network, and layer-by-layer edge weight information are encoded, and the encoding unit encodes layer structure information indicating a layer structure of the neural network, and a new layer flag indicating whether each of the layers to be encoded is a layer to be updated from a corresponding layer of a reference model, or a new layer.