Shadow Neural Network for Debugging Target Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Artificial neural networks face instability and debugging challenges due to resource constraints, limiting the ability to monitor and debug their activity effectively.

Innovation Solution

A shadow neural network is introduced to monitor and generate events based on the activity of a target neural network, allowing for more comprehensive debugging and logging without impacting the target network's behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a shadow neural network is introduced to monitor the target neural network, then debugging and monitoring capabilities are improved, but device complexity increases

Engineering Contradiction:
Improvedebugging capabilityVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent creates a shadow neural network that is a copy of the target neural network. This copy monitors the activity of the target network by receiving the same input data and producing corresponding output signals, enabling debugging and monitoring without modifying the original target network's structure or behavior.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The shadow neural network acts as an intermediary between the input data and the target neural network. It receives input data, processes it through its own neural network structure, and generates output signals that mirror the target network's activity, serving as a mediator for monitoring and debugging purposes.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If comprehensive monitoring of all neurons and synapses is implemented, then measurement precision is improved, but use of energy increases

Engineering Contradiction:
Improvemonitoring precisionVSAvoidenergy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

Instead of modifying the target network to enable comprehensive monitoring (which would consume additional energy), the patent creates a separate shadow network that copies the target network's architecture. This shadow network processes data independently and provides complete monitoring information without adding energy consumption to the target network's operation.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The monitoring function is segmented from the target neural network by creating a separate shadow network. This segmentation allows comprehensive monitoring of all neurons and synapses through the shadow network while the target network continues to operate independently without additional energy burden for monitoring purposes.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS9558442B2Monitoring neural networks with shadow networks
Publication Date: 2017.01.31 QUALCOMM INC
  • US9558442B2 patent drawing
  • US9558442B2 patent drawing
  • US9558442B2 patent drawing

AI summary

A method for generating an event includes monitoring a first neural network with a second neural network. The method also includes generating an event based on the monitoring. The event is generated at the second neural network. The event may be generated based on a spike received at the second network during the monitoring.