Electronic Device Data Generation Using Relative Information
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning technologies require a large amount of data for effective training, which is often not available in real-world scenarios, limiting their applicability to fields with limited data.
Innovation Solution
An electronic device that constructs data pairs from at least three data points, acquires relative information between these points, and learns a transformation function based on this information, allowing for data generation and improvement of task performance using only a small amount of data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning is applied to fields with limited data availability, then the applicability of deep learning is restricted, but the requirement for abundant data cannot be met
Solution Approach 1:
The patent generates synthetic data that copies and transforms existing data patterns through learned transformations. The system creates artificial data pairs by applying transformation functions to source data, effectively copying the essential characteristics while generating new variations that expand the available data quantity without requiring additional real-world data collection
Solution Approach 2:
The system changes the parameters and characteristics of existing data through learned transformation functions. By modifying data parameters (such as spatial transformations, color adjustments, or feature transformations), the system generates diverse data variations from limited original data, thereby increasing the effective data amount while maintaining the underlying data distribution patterns
2Quantity of substance
If transformation function is learned from very small data amount, then data overfitting problem occurs, but training on sufficient data is not possible
Solution Approach 1:
The system incorporates feedback mechanisms where the learned transformation functions are continuously evaluated and refined. The generated synthetic data is fed back into the training process to improve the transformation functions, creating an iterative cycle that prevents overfitting by constantly adjusting and validating the learned parameters against the limited available data
Solution Approach 2:
The system performs preliminary actions by pre-processing and preparing the limited available data in advance. It extracts meaningful features and patterns from the small dataset before generating synthetic data, ensuring that the transformation functions are learned from the most informative aspects of the data, thereby reducing overfitting risks
3Productivity
If existing deep learning methods are used without data generation capability, then task performance is limited by data availability, but adding data generation increases system complexity
Solution Approach 1:
The patent merges the data generation capability with the existing deep learning task processing system. The transformation function learning module and data generation components are integrated into the same system architecture, allowing the system to both generate synthetic data and perform deep learning tasks using a unified framework, thereby improving productivity without proportionally increasing complexity
Solution Approach 2:
The system achieves multi-functionality by enabling a single system to perform multiple operations: learning transformation functions from data, generating synthetic data pairs, and executing deep learning tasks. This universal design allows the same infrastructure to serve both data augmentation and task execution purposes, improving overall productivity while managing system complexity
Data Source
AI summary
Various embodiments provide an electronic device for generating data and improving task performance by using only a very small amount of data without prior knowledge of an associative domain and an operating method thereof. According to various embodiments, the electronic device may be configured to construct a plurality of data pairs from at least three data points, acquire relative information between the data points with respect to each of the data pairs, and learn a transformation function between the data points based on the relative information. According to various embodiments, the transformation function which can be learnt without a data overfitting problem even in a very small amount of data can be provided.