Distributed Model Learning via Centralized Parameter Aggregation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network training methods face challenges in efficiently distributing model learning across multiple devices, particularly in applications like mobile devices, where data refinement and model updates require coordinated yet loosely connected systems.

Innovation Solution

A distributed model learning approach where a central server receives labeled examples from user devices, generates a classifier based on statistical features, and updates models by aggregating parameter updates from devices, allowing for continuous learning and inference while minimizing bandwidth and memory usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If distributed model learning is implemented across multiple devices, then model refinement capability and scalability are improved, but system coordination complexity and communication overhead increase

Engineering Contradiction:
Improvemodel refinement capabilityVSAvoidsystem coordination complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

A central server acts as an intermediary coordinator that receives model parameter updates from multiple user devices, aggregates them using techniques like federated averaging, and distributes refined models back to devices. This mediator approach enables distributed learning without requiring complex peer-to-peer coordination between devices, thus improving model refinement capability while managing system complexity through centralized orchestration.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If continuous model updates are performed across distributed devices, then inference performance is improved, but bandwidth consumption and memory usage increase

Engineering Contradiction:
Improveinference performanceVSAvoidbandwidth consumption
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The system extracts and transmits only the essential model parameter updates (gradients or weight differences) rather than complete models or raw data. This extraction approach allows continuous model refinement and performance improvement while significantly reducing the bandwidth consumption associated with transmitting full model copies or raw training data across the distributed network.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Devices perform partial model updates locally and transmit only the necessary parameter changes to the central server, rather than performing complete retraining or transmitting excessive data. This partial action approach maintains inference performance through continuous refinement while optimizing bandwidth usage by sending only the minimal required updates.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If distributed training data is collected from multiple users, then model accuracy and generalization are improved, but data privacy risks and security concerns increase

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata privacy risks
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The central server acts as a trusted intermediary that aggregates model parameter updates from multiple users without directly accessing their private training data. This mediator architecture enables the system to leverage diverse user data for improved model accuracy and generalization while maintaining data privacy, as the server only receives encrypted or processed parameter updates rather than raw sensitive data.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

Instead of centralizing raw training data, the system copies and processes only the essential model parameter information across the distributed network. This copying approach allows multiple users to contribute to model improvement through their local data while preventing exposure of sensitive raw data, thus enhancing model accuracy without proportionally increasing privacy risks.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3143563B1Distributed model learning
Publication Date: 2023.05.24 QUALCOMM INC
  • EP3143563B1 patent drawingFigure 1
  • EP3143563B1 patent drawingFigure 2
  • EP3143563B1 patent drawingFigure 3

AI summary

A method of learning a model includes receiving model updates from one or more users. The method also includes computing an updated model based on a previous model and the model updates. The method further includes transmitting data related to a subset of the updated model to the a user(s) based on the updated model.