Distributed Neural Network Training via Parameter Exchange

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network models face challenges in training due to the need for large datasets, storage requirements, accessibility issues, and the difficulty in scaling laboratory-developed models to production environments, particularly in speech processing where data privacy and transmission constraints are significant.

Innovation Solution

A distributed system for training neural network models, where a master device and slave device collaborate to split training tasks, using a teacher-student configuration to exchange weight or gradient information over a network, maintaining privacy and infrastructure separation, and allowing training on unlabeled data to reduce labeling burdens.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large dataset is collected for accurate neural network predictions, then prediction accuracy is improved, but data storage requirements and transmission costs increase

Engineering Contradiction:
Improveprediction accuracyVSAvoiddata storage requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The training process is segmented across multiple devices (master device and slave devices). Each slave device trains a version of the neural network model locally using its own data, and only model parameters (weights and biases) are exchanged with the master device. This segmentation eliminates the need to centralize large datasets, reducing storage and transmission requirements while maintaining prediction accuracy through distributed learning.

Inventive Principle:
Principle #1Segmentation

2Ease of manufacture

If training data is transmitted between devices for centralized training, then model training is simplified, but data privacy and security are compromised

Engineering Contradiction:
Improvemodel training simplicityVSAvoiddata privacy risks
Core Design Contradiction:
Ease of manufactureVSObject-affected harmful factors

Solution Approach 1:

The invention extracts only the essential training information (model parameters such as weights and biases) from the training process and exchanges these between devices, while leaving the actual training data localized on each slave device. This extraction approach simplifies model training coordination while preserving data privacy, as the sensitive training data never leaves the local devices.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If comprehensive training data is made accessible to all training devices, then training efficiency is improved, but infrastructure complexity and access control difficulties increase

Engineering Contradiction:
Improvetraining efficiencyVSAvoidinfrastructure complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The master device serves multiple functions: it coordinates the distributed training process, aggregates model parameters from multiple slave devices, and manages the overall training workflow. This multi-functionality allows the system to achieve training efficiency comparable to centralized training without requiring complex data sharing infrastructure, as the master device handles all coordination through standardized parameter exchange.

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

4Reliability

If laboratory-developed neural network models are scaled to production environments, then real-world application capability is improved, but scalability and adaptability challenges arise

Engineering Contradiction:
Improvereal-world application capabilityVSAvoidscaling adaptability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The distributed training system is designed to be dynamic and adaptable: slave devices can be added or removed from the training network, different devices can contribute different amounts of data, and the master device dynamically adjusts the training coordination. This dynamic architecture allows laboratory-developed models to scale effectively to production environments without requiring rigid infrastructure, improving both reliability and adaptability.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3855365A1Distributed neural network model training
Publication Date: 2021.07.28 SOUNDHOUND INC
  • EP3855365A1 patent drawingFigure 1
  • EP3855365A1 patent drawingFigure 2
  • EP3855365A1 patent drawingFigure 3

AI summary

Systems and methods for distributed training of a neural network model are described. Various embodiments include a master device and a slave device. The master device has a first version of the neural network model. The slave device is communicatively coupled to a first data source and the master device, and the first data source is inaccessible by the master device, in accordance with one embodiment. The slave device is remote from the master device. The master device is configured to output first configuration data for the neural network model based on the first version of the neural network model. The slave device is configured to use the first configuration data to instantiate a second version of the neural network model. The slave device is configured to train the second version of the neural network model using data from the first data source and to output second configuration data for the neural network model. The master device is configured to use the second configuration data to update parameters for the first version of the neural network model.