Pseudo Data Generation for Transfer Learning Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Transfer learning in machine learning often fails when the transfer destination task significantly differs from the transfer source task, especially when there is limited variation and amount of training data, leading to low motion recognition accuracy.
Innovation Solution
A machine learning apparatus and method that generates pseudo data by processing real data from both tasks and uses this pseudo data for first transfer learning, followed by second transfer learning with real data to construct a second identification dictionary, thereby improving the success rate of transfer learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If transfer learning is performed from a transfer source task to a transfer destination task that is greatly different, then the learning result may not be useful for the transfer destination task, but performing relearning in the identifier is complex and time-consuming
Solution Approach 1:
The patent introduces pseudo data as an intermediary between the transfer source task and transfer destination task. The pseudo data generation unit creates synthetic training data that bridges the gap between different tasks, allowing the identifier to gradually adapt from the source task to the destination task through intermediate pseudo data, rather than direct transfer or complete relearning
Solution Approach 2:
The patent performs preliminary action by generating pseudo data before actual transfer learning to the destination task. The pseudo data is prepared in advance to facilitate smoother transfer, and the system also performs preliminary evaluation of task similarity before initiating transfer learning
2Adaptability or versatility
If an identifier is caused to learn a task for which a large amount of training data cannot be prepared, then the variety of motions that can be identified is limited, but preparing a large amount of training data with wide variation is difficult for rare motions
Solution Approach 1:
The patent uses copying by generating pseudo data that replicates the characteristics of real training data. The pseudo data generation unit creates synthetic copies of training samples with varied transformations (rotations, scaling, translations, etc.), providing sufficient quantity and diversity of training data without requiring actual collection of rare motion data
Solution Approach 2:
The patent applies parameter changes by transforming the pseudo data through various parameter modifications including rotation angles, scaling factors, translation distances, and other geometric transformations. These parameter changes create diverse variations from limited source data, expanding the effective training dataset
3Reliability
If the variation and amount of training data in a transfer destination task is too small, then the transfer is likely to fail, but generating sufficient varied data requires complex data augmentation
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate its own training data through pseudo data generation. The identifier and pseudo data generation unit work together to create appropriate training samples without requiring external data collection or complex manual data augmentation processes
Data Source
AI summary
A machine learning apparatus 100 is an apparatus for constructing, by transfer learning, a second identification dictionary to be used in a second task from a first identification dictionary to be used in a first task. The machine learning apparatus 100 includes: a pseudo data generation unit 10 that generates pseudo data by processing one of real data in the first task and real data in the second task or both pieces of real data; and a learning unit 20 that constructs a third identification dictionary by performing, using the first identification dictionary, first transfer learning using the pseudo data as training data, and furthermore, constructs the second identification dictionary by performing, using the third identification dictionary, second transfer learning using the real data in the second task as training data.