Distributed Neural Network Training for Private Remote Event Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network training and validation methods face challenges in maintaining data privacy and efficiency, particularly when training and validating models on remote servers, as they often require transmitting sensitive data to a central server, which can compromise privacy and increase workload.
Innovation Solution
A distributed deep learning approach where neural networks are trained and validated on multiple remote servers using data private to each server, with updates and evaluations performed locally, ensuring data remains private and reducing the central server's workload by adaptively filtering and customizing models based on local data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data is transmitted to a central server for training and validation, then model training and validation can be performed centrally, but data privacy is compromised and workload on the central server increases
Solution Approach 1:
The system segments the centralized training process into distributed components across multiple servers. Each server independently trains and validates neural network models using local data, eliminating the need to transmit sensitive data to a central server while distributing the computational workload.
Solution Approach 2:
The patent introduces an intermediary mechanism where servers exchange only model parameters and validation results rather than raw data. This intermediary approach allows collaborative model improvement while maintaining data privacy, as the actual sensitive data never leaves its originating server.
2Ease of manufacture
If data is transmitted to a central server for training, then centralized model training can be performed, but data transmission requirements and network bandwidth increase
Solution Approach 1:
The system extracts only the essential model parameters and validation metrics from the training process, transmitting only these compressed representations between servers rather than the full datasets. This dramatically reduces network bandwidth requirements while preserving the ability to perform centralized model aggregation.
3Loss of information
If neural networks are trained on local data without transmission, then data privacy is maintained, but model accuracy and customization may be limited
Solution Approach 1:
The system merges models trained on local data across multiple servers through parameter aggregation and ensemble methods. This combining approach allows each server to leverage its local data privacy while collectively achieving higher model accuracy by incorporating diverse local knowledge from multiple sources.
Solution Approach 2:
The patent implements feedback mechanisms where validation results from local models are exchanged between servers. This feedback loop allows servers to refine their local models based on performance metrics from other servers, improving overall model accuracy without compromising data privacy or requiring raw data transmission.
Data Source
AI summary
Apparatus and associated methods relate to training a neural network on a first host system, sending the neural network to a second host system, training the neural network by the second host system based on data private to the second host system, and employing the neural network to filter events sent to the first host system. In an illustrative example, the first host system may be a server having a central repository including trained neural networks and historical data, and the second host system may be a remote server having a data source private to the remote server. The private remote data source may be a camera. In some examples, events may be filtered as a function of a prediction of error in the neural network. Various examples may advantageously provide remote intelligent filtering. For example, remote data may remain private while adaptively filtering events to the central server.