Learned-Model Intermediate Representation Control for Valid Outputs

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing deep learning models, such as the Transformer model and AlphaFold2, lack clarity in determining the manipulation range of intermediate representations to generate valid outputs, making it difficult to produce high-quality outputs without altering model parameters.

Innovation Solution

A control method that evaluates the validity of current outputs using indicators correlated with the probability distribution of learning data and adjusts intermediate representations to minimize an energy function, ensuring the generation of valid outputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the intermediate representation is manipulated to generate diverse outputs, then the versatility of the learned model increases, but the reliability of output data decreases due to unclear manipulation range

Engineering Contradiction:

Improveversatility of learned modelVSAvoidreliability of output data

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent introduces a feedback mechanism where the validity of output data is evaluated based on the data distribution of learning data, and this evaluation result is fed back to guide the manipulation of intermediate representation. The control device adjusts the intermediate representation iteratively to maintain output validity, creating a closed-loop system that ensures reliability while enabling versatility.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent changes the parameters of intermediate representation within a controlled range determined by data distribution analysis. By identifying the valid manipulation range through statistical analysis of learning data, the system allows parameter changes that generate diverse outputs while staying within bounds that guarantee output reliability.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If the manipulation range of intermediate representation is expanded to generate various outputs, then the productivity increases, but the manufacturing precision decreases due to loss of validity control

Engineering Contradiction:

Improveoutput generation efficiencyVSAvoidvalidity of output data

Core Design Contradiction:

ProductivityVSManufacturing precision

Solution Approach 1:

The patent performs preliminary analysis of the data distribution of learning data before manipulating intermediate representation. By pre-determining the valid manipulation range through statistical analysis, the system prepares constraints that guide subsequent output generation, ensuring both efficiency and validity without requiring iterative validation of each output.

Inventive Principle:

Principle #10Preliminary action

3Loss of information

If the parameters of learned model are kept fixed to utilize representation capability, then the loss of information is reduced, but the ease of operation decreases due to inability to adjust model behavior

Engineering Contradiction:

Improvepreservation of learned knowledgeVSAvoidcontrollability of model output

Core Design Contradiction:

Loss of informationVSEase of operation

Solution Approach 1:

The patent introduces intermediate representation as a mediator between the fixed learned model and the desired outputs. By manipulating this intermediate layer rather than model parameters, the system enables flexible control of model behavior while preserving the learned knowledge encapsulated in the fixed parameters. The intermediate representation serves as a controllable interface that decouples model integrity from operational flexibility.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20250292090A1Computer-readable recording medium storing control program, control method, and information processing device

Publication Date: 2025.09.18 FUJITSU LTD

AI summary

A non-transitory computer-readable recording medium storing a control program for causing a computer to execute a process includes, when an intermediate representation of input data input to a learned model is changed, evaluating validity of current output data generated from the intermediate representation by the learned model based on a first indicator correlated with an existence probability of output data in a data distribution of learning data used for learning of the learned model, and changing the intermediate representation based on an evaluation result such that validity of output data generated by the learned model increases.