Dynamic Reconfiguration Training Architecture for Neural Networks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Configuring machine learning architectures for specific tasks is cumbersome and often requires modifications to code or hardware, leading to difficulties in tracking changes and reproducing results due to lost parameters and configuration settings.

Innovation Solution

A dynamic reconfiguration training computer architecture (DRTA) that uses a configuration file with a tree structure to specify parameters for machine learning sessions, allowing for easy tracking and reproduction of neural network architectures and results, implemented in a JSON file format for ease of use.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If machine learning architecture is configured for a specific task using traditional methods, then the architecture can be customized for the task, but the configuration process becomes cumbersome and requires modification to code or hardware

Engineering Contradiction:

Improvetask-specific customizationVSAvoidconfiguration process

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent uses configuration files that serve as templates or copies of the desired machine learning architecture settings. Instead of modifying code or hardware directly, users create configuration files that replicate the intended architecture parameters, making the customization process simpler and more reproducible

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent enables task-specific customization by allowing users to modify parameters in configuration files rather than changing code or hardware. This approach maintains adaptability while simplifying the configuration process, as users can adjust architectural parameters through file editing without complex development work

Inventive Principle:

Principle #35Parameter changes

2Productivity

If traditional configuration methods are used, then code or hardware modifications enable task-specific optimization, but tracking changes and reproducing results becomes difficult due to lost parameters and configuration settings

Engineering Contradiction:

Improvetask-specific optimizationVSAvoidconfiguration settings

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The patent performs preliminary action by automatically recording and storing all configuration parameters and settings in a structured configuration file before the machine learning training begins. This ensures that all necessary information is captured upfront, preventing loss of configuration data and enabling easy reproduction of results

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system provides feedback by maintaining a complete record of configuration settings that can be retrieved and used to reproduce results. This feedback mechanism ensures that no configuration information is lost, as all parameters are systematically stored and can be reviewed or reused

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS11551026B2Dynamic reconfiguration training computer architecture

Publication Date: 2023.01.10 RAYTHEON CO

AI summary

A dynamic reconfiguration training machine learning computer architecture is disclosed. According to some aspects, a computing machine accesses a configuration file. The configuration file specifies parameters for a machine learning session. The computing machine trains a machine learning module to solve a problem, where the machine learning module operates according to the parameters specified in the configuration file. The computing machine generates an output representing the trained machine learning module.