Selective Dataset Augmentation for Deep Learning Variance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated systems for generating training datasets for deep-learning models lack efficiency, as they augment all data dimensions without regard to necessary variance, leading to excessive processing requirements and time, while failing to ensure sufficient diversity across the dataset, resulting in inaccurate predictions when faced with variations.
Innovation Solution
A method and system that identify specific aspects of the training dataset with low variance and selectively augment those dimensions to achieve the predetermined level of variability, reducing processing resources and time by only augmenting data that needs it, thereby ensuring sufficient diversity and robustness of the dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If all data dimensions are augmented without selective analysis, then the dataset diversity is increased, but the processing time and computational resources are excessively consumed
Solution Approach 1:
The patent segments the data augmentation process by identifying and separating dimensions with low variance from those with sufficient variance. Only the low-variance dimensions are selected for augmentation, dividing the original comprehensive augmentation task into targeted sub-tasks that reduce overall processing time while maintaining necessary dataset diversity.
Solution Approach 2:
The patent applies local quality by treating different data dimensions differently based on their individual variance characteristics. Instead of uniformly augmenting all dimensions, the system selectively applies augmentation only to dimensions that benefit from it (those with variance below the threshold), optimizing resource allocation across different parts of the dataset.
2Adaptability or versatility
If all data dimensions are augmented without selective analysis, then the dataset diversity is increased, but the computational resources are excessively consumed
Solution Approach 1:
The patent segments the data augmentation process by identifying and separating dimensions with low variance from those with sufficient variance. Only the low-variance dimensions are selected for augmentation, dividing the original comprehensive augmentation task into targeted sub-tasks that reduce overall processing time while maintaining necessary dataset diversity.
Solution Approach 2:
The patent applies local quality by treating different data dimensions differently based on their individual variance characteristics. Instead of uniformly augmenting all dimensions, the system selectively applies augmentation only to dimensions that benefit from it (those with variance below the threshold), optimizing resource allocation across different parts of the dataset.
3Productivity
If selective augmentation is performed only on low-variance dimensions, then the processing efficiency is improved, but insufficient variance in some dimensions may remain
Solution Approach 1:
The patent implements feedback by calculating the variance of each dimension before augmentation and using this measurement to guide the augmentation process. Dimensions with variance below the threshold are identified and targeted for augmentation, creating a feedback loop that ensures variability is achieved where needed while avoiding unnecessary processing where variability already exists.
4Ease of manufacture
If traditional augmentation methods are used without variance analysis, then the process is simpler to implement, but the dataset may lack sufficient diversity for accurate predictions
Solution Approach 1:
The patent implements feedback by calculating the variance of each dimension before augmentation and using this measurement to guide the augmentation process. Dimensions with variance below the threshold are identified and targeted for augmentation, creating a feedback loop that ensures variability is achieved where needed while avoiding unnecessary processing where variability already exists.
Data Source
AI summary
One embodiment provides a method, including: receiving a training dataset to be utilized for training a deep-learning model; identifying a plurality of aspects of the training dataset, wherein each of the plurality of aspects corresponds to one of a plurality of categories of operations that can be performed on the training dataset; measuring, for each of the plurality of aspects, an amount of variance of the aspect within the training dataset; creating additional data to be incorporated into the training dataset, wherein the additional data comprise data generated for each of the aspects having a variance less than a predetermined amount, wherein the data generated for an aspect results in the corresponding aspect having an amount of variance at least equal to the predetermined amount; and incorporating the additional data into the training dataset.