Distributed Neural Network Layer Partitioning for Device Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional neural network structure search technologies are inadequate for distributed processing across multiple devices, as they focus on single-device calculations, neglecting the potential for optimizing network structures in distributed systems.

Innovation Solution

An information processing method that evaluates and determines the neural network structure by considering information transfer between a first device with low calculation performance (e.g., IoT devices) and a second device with high calculation performance (e.g., cloud servers), using genetic operations to find an optimal structure that balances recognition performance, calculation amount, and energy savings.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional single-device neural network structure search is used, then calculation performance is optimized for individual devices, but distributed processing efficiency and recognition performance are not improved

Engineering Contradiction:

Improvedistributed processing efficiencyVSAvoidrecognition performance

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The neural network structure is segmented into multiple layers that are distributed across different devices. The evaluation unit divides the neural network into first layers processed by a first device and second layers processed by a second device, enabling distributed processing while maintaining overall system functionality and recognition performance.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The invention changes the evaluation parameters to include not only calculation amount but also information transfer characteristics between devices. By evaluating based on transfer information characteristics and adjusting network structure according to these distributed processing parameters, the system optimizes both distributed efficiency and recognition performance simultaneously.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If network structure is optimized for single-device calculation, then calculation amount is reduced, but information transfer efficiency in distributed systems is not considered

Engineering Contradiction:

Improvecalculation efficiencyVSAvoidcommunication load

Core Design Contradiction:

ProductivityVSLoss of energy

Solution Approach 1:

The evaluation parameters are changed to include information transfer characteristics between devices in addition to calculation amount. This allows the network structure to be optimized for both calculation efficiency and communication efficiency in distributed systems, reducing overall energy loss.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The evaluation unit provides feedback on the overall system performance including both calculation and communication aspects. This feedback mechanism allows iterative optimization of the network structure to balance calculation efficiency with communication load reduction across distributed devices.

Inventive Principle:

Principle #23Feedback

3Ease of operation

If neural network structure is simplified for low-performance devices, then ease of operation is improved, but recognition performance deteriorates

Engineering Contradiction:

Improvedevice compatibilityVSAvoidrecognition accuracy

Core Design Contradiction:

Ease of operationVSMeasurement precision

Solution Approach 1:

The neural network is segmented into multiple layers distributed across devices with different performance capabilities. Low-performance devices handle simplified first layers while higher-performance devices handle more complex second layers, maintaining ease of operation across all devices while preserving overall recognition accuracy through the complete network structure.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different parts of the neural network have different structural qualities optimized for their specific device environments. Each device processes layers with complexity appropriate to its computational capabilities, allowing local optimization while maintaining global recognition performance through the integrated network.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS12462160B2Performing distributed processing using layers of a neural network divided between a first device and a second device

Publication Date: 2025.11.04 SONY GROUP CORP

AI summary

An information processing method according to the present disclosure includes the steps of: evaluating, by a computer, a neural network having a structure held in a divided manner by a first device and a second device based on information on transfer of information between the first device and the second device in the neural network; and determining, by the computer, the structure of the neural network based on the evaluation of the neural network.