Compressed Training Data Sharing for Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for training neural networks face challenges such as confidentiality and privacy issues, especially in computer vision applications, as well as bandwidth and cost constraints due to the need to process and transfer high-resolution images or videos for object detection tasks.
Innovation Solution
An information processing device and method for sharing compressed training data, which extracts intermediate results from neural networks deployed in live environments, reducing the computational load and privacy concerns by processing and transmitting feature maps from selected intermediate layers instead of full images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images or videos are collected and transferred for neural network training, then training data quality is improved, but bandwidth consumption increases and privacy issues arise
Solution Approach 1:
The patent extracts intermediate feature maps from deployed neural networks instead of transferring original high-resolution images. This extraction approach obtains essential training information while eliminating unnecessary pixel data, thus improving training data quality without proportionally increasing bandwidth consumption.
Solution Approach 2:
The patent creates compressed representations (feature maps) as copies of the essential information from original images. These feature map copies contain the necessary training patterns but occupy significantly less bandwidth space than the original high-resolution images.
2Measurement precision
If high-resolution images or videos are processed on training servers, then model accuracy is improved, but computational costs increase
Solution Approach 1:
The patent performs preliminary processing by extracting feature maps at the edge device where the neural network is already deployed. This preliminary extraction prepares the data in an optimized format before transfer, reducing the computational burden on training servers while maintaining model accuracy.
3Measurement precision
If original images are transferred for training, then detection accuracy is improved, but privacy and confidentiality issues worsen
Solution Approach 1:
The patent extracts only the essential feature representations from images, leaving out personally identifiable information and sensitive details. This extraction maintains detection accuracy by preserving key patterns while eliminating privacy concerns associated with transferring original images.
Data Source
AI summary
An information processing device and method for sharing of compressed training data for neural network training is provided. The information processing device receives a first image which includes an object of interest. The information processing device extracts, from the received first image, a region of interest which includes the object of interest. Once extracted, the extracted region of interest is provided to an input layer of N numbers of layers of a first neural network, trained on an object detection task. The information processing device selects an intermediate layer of the first neural network and extracts a first intermediate result as an output generated by the selected intermediate layer of the first neural network based on the input RoI. Once extracted, the information processing device shares the extracted first intermediate result as compressed training data with a server to train a second neural network on the object detection task.