Federated Embedding Slices for Low-Bandwidth Private Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning algorithms for learning entity embeddings are infeasible in distributed environments with unreliable network connections and low computational power, particularly when dealing with extremely large vocabularies of entities, as they require excessive communication and compromise user privacy.

Innovation Solution

A slice-based embedding serving scheme in federated learning, where clients receive only smaller subsets of embeddings organized by semantic relatedness, allowing local updates while preserving privacy through obfuscation and reducing communication overhead.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If clients receive the complete vocabulary of embeddings for training, then the model can learn from all entities, but the communication overhead becomes excessive and privacy is compromised

Engineering Contradiction:
Improveembedding qualityVSAvoidcommunication overhead
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The vocabulary of embeddings is divided into multiple slices, where each slice contains a subset of embeddings organized by semantic relatedness. Clients receive only the relevant slice containing embeddings for entities in their local dataset, rather than the complete vocabulary. This segmentation reduces communication overhead while maintaining the ability to learn high-quality embeddings for local entities through federated averaging across multiple clients.

Inventive Principle:
Principle #1Segmentation

2Loss of energy

If clients receive only a subset of embeddings, then communication costs are reduced, but the ability to learn from diverse entities is limited

Engineering Contradiction:
Improvecommunication costVSAvoidlearning capability
Core Design Contradiction:
Loss of energyVSAdaptability or versatility

Solution Approach 1:

The vocabulary is segmented into multiple slices organized by semantic relatedness, allowing clients to receive targeted subsets of embeddings relevant to their local data while maintaining learning capability for those specific entities.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Embedding updates from multiple clients are combined through federated averaging to create a global embedding model. This merging process allows the system to learn from diverse entities across all clients while each client only processes its local subset, thus maintaining versatility without increasing individual communication costs.

Inventive Principle:
Principle #5Merging (Combining)

3Measurement precision

If the complete vocabulary of embeddings is transmitted to clients, then all entities can be represented, but user privacy is compromised

Engineering Contradiction:
Improveentity representationVSAvoidprivacy loss
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The embedding vocabulary is segmented into slices, and clients receive only the slice containing embeddings for entities in their local dataset. This selective transmission minimizes the exposure of embedding data, thereby reducing privacy loss while still providing accurate entity representations for local training.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system extracts and transmits only the necessary subset of embeddings (those corresponding to entities in the client's local dataset) rather than the complete vocabulary. This extraction approach maintains entity representation quality for local entities while removing unnecessary embedding data that would contribute to privacy loss.

Inventive Principle:
Principle #2Taking out (Extraction)

4Extent of automation

If federated learning is implemented with large vocabulary embeddings, then distributed training is enabled, but computational requirements exceed client capabilities

Engineering Contradiction:
Improvedistributed trainingVSAvoidcomputational power
Core Design Contradiction:
Extent of automationVSPower

Solution Approach 1:

The large vocabulary embedding task is segmented into smaller slices, reducing the computational burden on each client. Clients only need to process and update embeddings for entities in their local dataset rather than handling the entire vocabulary, making distributed training feasible on resource-constrained devices.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Clients perform partial training by updating only the subset of embeddings relevant to their local data rather than training on the complete vocabulary. This partial action approach enables distributed training on clients with limited computational power while still contributing to the global model through federated averaging.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4049162B1Online federated learning of embeddings
Publication Date: 2025.12.03 GOOGLE LLC
  • EP4049162B1 patent drawingFigure 1
  • EP4049162B1 patent drawingFigure 2
  • EP4049162B1 patent drawingFigure 3A

AI summary

The present disclosure provides for the generation of embeddings within a machine learning framework, such as, for example, a federated learning framework in which a high-quality centralized model is trained on training data distributed over a large number of clients each with unreliable network connections and low computational power. In an example federated learning setting, in each of a plurality of rounds, each client independently updates the model based on its local data and communicates the updated model back to the server, where all the client-side updates are used to update a global model. The present disclosure provides systems and methods that may generate embeddings with local training data while preserving the privacy of a user of the client device.