Proxy Prediction Models for On-Device Machine Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine learning methods face challenges in achieving accurate and personalized predictions on client devices while maintaining user privacy, as they either require intrusive calibration or collect and use crowd-sourced data that may violate privacy.

Innovation Solution

A system where a server generates and provides prediction models to client devices, which select and adapt a proxy model based on private user data to enhance user experience without compromising privacy, using divergence estimation to match user data distributions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If machine learning is trained using only on-device user data, then user privacy is preserved, but training accuracy and time are degraded

Engineering Contradiction:
Improveprivacy protectionVSAvoidprediction accuracy
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The system segments the training process into two distinct phases: (1) server-side training using crowd-sourced data to create initial prediction models, and (2) client-side fine-tuning using local user data to personalize the model. This segmentation allows each phase to optimize for its specific goal while maintaining overall system effectiveness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The server performs preliminary training actions by collecting crowd-sourced data and generating base prediction models before they are deployed to client devices. This preliminary action provides a head start for on-device training, reducing the amount of local data needed and accelerating the personalization process while maintaining privacy.

Inventive Principle:
Principle #10Preliminary action

2Manufacturing precision

If machine learning models are trained using crowd-sourced data on a server, then prediction accuracy improves, but user privacy may be violated

Engineering Contradiction:
Improveprediction accuracyVSAvoidprivacy protection
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The system extracts only the essential training patterns and knowledge from crowd-sourced data on the server, then removes all personally identifiable information. Only the distilled prediction model parameters are transferred to client devices, leaving no sensitive user data on the server.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The server creates copies of the prediction model trained on crowd-sourced data and distributes these model copies to client devices. The actual user data remains on the server in aggregated form, while each device receives a personalized model copy that can be further fine-tuned locally without sharing sensitive information.

Inventive Principle:
Principle #26Copying

3Manufacturing precision

If sensor calibration is performed to personalize machine learning for individual users, then prediction accuracy improves, but the calibration process becomes intrusive and time-consuming

Engineering Contradiction:
Improvepersonalization accuracyVSAvoidcalibration convenience
Core Design Contradiction:
Manufacturing precisionVSEase of operation

Solution Approach 1:

The system performs the time-consuming calibration process in advance on the server using crowd-sourced data from multiple users. The pre-calibrated model parameters are then transferred to individual devices, eliminating the need for users to undergo intrusive calibration procedures while still achieving personalized accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The server performs universal calibration by collecting data from diverse users and creating generalized calibration parameters that work across multiple individuals. This multi-functional approach allows a single calibration process to benefit many users simultaneously, reducing the need for individual calibration sessions.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3461293B1Using proxies to enable on-device machine learning
Publication Date: 2024.03.20 APPLE INC
  • EP3461293B1 patent drawingFigure 1
  • EP3461293B1 patent drawingFigure 2
  • EP3461293B1 patent drawingFigure 3

AI summary

Systems and methods are disclosed for a server generating a plurality of proxy prediction models for each of a plurality of applications for use by client devices. Providing a prediction model based on a data distribution that more closely matches the user private data of the actual user of the client device, as opposed to a generic model, can improve the accuracy of predictions and improve the benefits to the user of the client device. The plurality of prediction models are generated by a server and provided to a plurality of clients. For a particular application, a client device can use previously collected private user data to select one of the plurality of prediction models for the application as being a closest matching prediction model to use for a particular user.