Generative Model Training With Distance-Marked Labels for Concise Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative models often generate outputs that are too long and inaccurate, making them difficult to comprehend and prone to errors.
Innovation Solution
A method of training a generative model by inserting markers at specific points in an input sequence to control output length, using markers that indicate the distance to the end point, and adjusting model parameters based on these markers and a preset rule to generate accurate and concise outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If the generative model generates detailed outputs, then the completeness of information is improved, but the output length increases and accuracy decreases
Solution Approach 1:
The patent segments the label sequence by inserting multiple markers at different positions (beginning, middle, end) to divide the generation process into controlled segments. This allows the model to generate complete information while maintaining accuracy through structured segmentation of the output sequence.
Solution Approach 2:
The patent introduces markers as intermediary elements within the label sequence. These markers serve as mediators that guide the generative model to produce accurate outputs by providing structural cues at specific positions, thereby resolving the conflict between completeness and precision.
2Loss of information
If the generative model generates comprehensive outputs, then the information coverage is improved, but the output becomes too long and difficult to comprehend
Solution Approach 1:
By segmenting the label sequence with strategically placed markers, the patent enables the model to generate comprehensive information in a structured format that avoids excessive length. The segmentation allows efficient information delivery without unnecessary expansion.
Solution Approach 2:
The markers are inserted preliminarily into the label sequence before generation, establishing a predetermined structure that guides the model to produce comprehensive yet concise outputs. This preliminary structuring prevents the generation of overly long sequences while maintaining information coverage.
3Manufacturing precision
If markers are inserted at multiple points in the label sequence, then the output length control is improved, but the training complexity increases
Solution Approach 1:
The use of multiple markers segments the training process into manageable parts, each marker providing specific length control cues. This segmentation improves output length control while the modular nature of marker insertion keeps training complexity manageable through systematic processing.
Data Source
AI summary
A method, performed by an electronic device, of training a generative model, the method including: obtaining a first label for an input sequence; generating a second label from the first label using a plurality of markers comprising information on a distance between a respective marker from the plurality of markers and an end point of the first label; training the generative model based on the input sequence and the second label; and modifying one or more parameters of the generative model based on the training of the generative model.