Local Data Evaluation for Private Machine-Learned Model Selection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge of managing and evaluating machine-learned models on user devices is hindered by the need for centralized data logging, which raises privacy concerns and bandwidth issues, and the unpredictability of data utility, leading to inefficiencies in model training and personalization.

Innovation Solution

Implementing a system where user devices manage and evaluate machine-learned models using locally logged data, evaluating performance metrics to activate models based on local data, and selectively uploading minimal necessary information for training or aggregation, while ensuring user privacy through local storage and differential privacy techniques.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If data is uploaded centrally to train and evaluate machine-learned models, then model training and evaluation can be performed, but user privacy is compromised and bandwidth requirements increase

Engineering Contradiction:
Improvemodel training qualityVSAvoiduser privacy risk
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

Instead of uploading data to servers for model training and evaluation, the patent inverts the approach by downloading pre-trained models to user devices and evaluating them locally using locally stored data. This reversal eliminates the need to centralize sensitive user data while maintaining model functionality through local inference and evaluation.

Inventive Principle:
Principle #13The other way round (Inversion)

Solution Approach 2:

The patent extracts the model evaluation function from the centralized server environment and places it at the edge device. By extracting only the necessary model artifacts and evaluation capabilities to user devices, the system maintains model quality assessment without requiring centralization of data or computation.

Inventive Principle:
Principle #2Taking out (Extraction)

2Quantity of substance

If all user data is logged centrally, then sufficient data for model training is available, but bandwidth and power requirements become infeasible

Engineering Contradiction:
Improvedata volume for trainingVSAvoidbandwidth and power consumption
Core Design Contradiction:
Quantity of substanceVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the essential model files and evaluation metrics from the centralized system and places them at edge devices. This extraction eliminates the need to transmit and store large volumes of raw user data centrally, reducing bandwidth and power requirements while maintaining model functionality through local processing.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of centralizing data collection and processing, the patent inverts the architecture by distributing model artifacts to edge devices where local data already exists. This inversion eliminates the need for continuous data uploads and central processing, dramatically reducing bandwidth and power consumption.

Inventive Principle:
Principle #13The other way round (Inversion)

3Object-affected harmful factors

If data is logged locally at user devices, then user privacy is protected and bandwidth is reduced, but model training and evaluation capabilities are limited

Engineering Contradiction:
Improveuser privacy protectionVSAvoidmodel training capability
Core Design Contradiction:
Object-affected harmful factorsVSAdaptability or versatility

Solution Approach 1:

The patent applies preliminary action by pre-training models centrally using aggregated data from multiple users, then distributing these pre-trained models to edge devices. This preliminary training phase ensures models have robust general capabilities before deployment, while local evaluation fine-tunes them to specific user contexts without requiring continuous central data collection.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the model lifecycle into distinct phases: centralized pre-training using aggregated data, local deployment with privacy protection, and local evaluation using locally stored data. This segmentation allows each phase to operate in its optimal environment - centralized for broad learning, distributed for privacy-preserving personalization.

Inventive Principle:
Principle #1Segmentation

4Measurement precision

If multiple machine-learned models are evaluated locally, then informed model selection is possible, but device computational resources are consumed

Engineering Contradiction:
Improveperformance metric accuracyVSAvoiddevice computational power
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent applies partial action by evaluating only the most critical performance metrics locally rather than conducting exhaustive model assessments. This selective evaluation approach provides sufficient information for informed model selection while minimizing computational overhead on resource-constrained edge devices.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3944160B1Management and evaluation of machine-learned models based on locally logged data
Publication Date: 2025.08.06 GOOGLE LLC
  • EP3944160B1 patent drawingFigure 1A
  • EP3944160B1 patent drawingFigure 1B
  • EP3944160B1 patent drawingFigure 1C

AI summary

The present disclosure provides systems and methods for the management and/or evaluation of machine-learned models based on locally logged data. In one example, a user computing device can obtain a machine-learned model (e.g., from a server computing device) and can evaluate at least one performance metric for the machine-learned model. In particular, the at least one performance metric for the machine-learned model can be evaluated relative to data that is stored locally at the user computing device. The user computing device and/or the server computing device can determine whether to activate the machine-learned model on the user computing device based at least in part on the at least one performance metric. In another example, the user computing device can evaluate a plurality of machine-learned models against locally stored data. At least one of the models can be selected based on the evaluated performance metrics.