Compressed Neural Network Transmission for IoT Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network technologies face challenges in efficient, error-resilient, and secure transmission and processing due to high computational complexity and battery consumption, particularly in IoT devices, which necessitate distributed computation scenarios and standardized communication formats.

Innovation Solution

The development of a method and apparatus for neural networks that enable efficient, error-resilient, and secure transmission and reception of compressed neural networks between clients and servers, utilizing standardized communication and representation formats to support intelligent and flexible sharing of neural networks as web-based services in cloud environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If neural networks are transmitted and processed in centralized cloud servers, then computational accuracy and model performance are improved, but data transmission latency and network dependency increase

Engineering Contradiction:
Improvecomputational accuracyVSAvoiddata transmission latency
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the neural network into compressed core weights and differential update weights, allowing selective transmission and processing. The compressed representation enables local device processing while maintaining model performance, reducing latency by avoiding full model transmission.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a compression intermediary layer that transforms full-precision neural network weights into compressed representations with differential updates. This intermediary mechanism enables efficient local processing while maintaining connection to cloud-based model improvements.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If full-precision neural networks are deployed on IoT devices, then model performance is improved, but device computational burden and energy consumption increase

Engineering Contradiction:
Improvemodel performanceVSAvoiddevice energy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the parameter representation from full-precision weights to compressed differential representations. By transforming the weight parameters into a compressed format with incremental updates, the system reduces memory bandwidth requirements and computational energy while maintaining model performance.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses lightweight compressed weight representations that can be quickly updated and discarded in favor of newer compressed versions. The differential update mechanism allows efficient replacement of weight representations without requiring full model retransmission.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Productivity

If neural network weights are compressed for efficient transmission, then transmission efficiency is improved, but weight precision and reconstruction accuracy deteriorate

Engineering Contradiction:
Improvetransmission efficiencyVSAvoidweight precision
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent performs preliminary compression of neural network weights into a compact representation before transmission. By pre-computing the compressed form and maintaining differential update capabilities, the system prepares the model for efficient transmission while preserving reconstruction accuracy through incremental updates.

Inventive Principle:
Principle #10Preliminary action

4Adaptability or versatility

If standardized communication formats are implemented for neural networks, then interoperability and sharing capability are improved, but system complexity and standardization overhead increase

Engineering Contradiction:
ImproveinteroperabilityVSAvoidstandardization overhead
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements a universal compressed neural network representation format that can be used across different devices and platforms. The standardized compression and differential update mechanism enables broad interoperability while maintaining implementation efficiency through a unified approach.

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

Data Source

PatentEP3683733B1A method, an apparatus and a computer program product for neural networks
Publication Date: 2024.11.06 NOKIA TECHNOLOGIES OY
  • EP3683733B1 patent drawingFigure 1~2
  • EP3683733B1 patent drawingFigure 3
  • EP3683733B1 patent drawingFigure 4

AI summary

The invention relates to a method, comprising generating a request for a neural network, wherein the generating comprises including into the request information on at least one requirement for the neural network; sending the request to a server; receiving from the server a response comprising at least means for identifying a compressed version of the requested neural network, said neural network having been compressed according to the at least one requirement; and causing the neural network to perform a task, wherein the neural network is accessed by means of said identification.