Domain Adaptation via Pseudo-Sample Generation for Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Domain adaptation techniques face accuracy issues when the target domain lacks samples corresponding to all labels from the source domain, leading to deteriorated inference accuracy in machine learning models.
Innovation Solution
An information processing apparatus and method that generates pseudo-samples in the target domain based on the distribution of samples from the source domain, using confidence scores and data transformation to approximate the distribution of both domains in the feature space, thereby supplementing the target domain and improving domain adaptation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If domain adaptation is performed using conventional methods, then the learned model can be applied to the target domain, but the accuracy deteriorates when the target domain lacks samples corresponding to all labels from the source domain
Solution Approach 1:
The patent generates pseudo-samples by copying the distribution characteristics of available target domain samples to create synthetic samples for missing classes. Specifically, it uses the distribution of samples from the first class in the target domain to generate pseudo-samples of the second class, effectively copying statistical properties to fill gaps in the target domain data
Solution Approach 2:
The patent performs preliminary generation of pseudo-samples before the actual domain adaptation training. By pre-generating pseudo-samples that approximate the missing target domain distribution, the system prepares the necessary training data in advance, allowing the learned model to be trained on a more complete and balanced dataset that includes both real and synthetic target domain samples
2Manufacturing precision
If the target domain samples are insufficient for all source domain labels, then the domain adaptation cannot sufficiently approximate the sample distribution, but increasing target domain data collection is not always feasible
Solution Approach 1:
The patent copies the distribution pattern of available target domain samples (first class) to generate pseudo-samples for the missing class (second class). This allows the system to approximate the target domain distribution for all source domain labels without requiring actual target domain samples for every label, thereby achieving high distribution approximation accuracy with limited target domain data
Solution Approach 2:
The patent transforms the feature space representation of available target domain samples to generate pseudo-samples. By applying data transformation in the feature space and using machine learning to approximate distributions, the system changes the parameters and characteristics of existing samples to create new synthetic samples that fill the gaps in target domain coverage
Data Source
AI summary
Disclosed herein is an information processing apparatus, comprising: a feature extraction unit configured to extract features from a sample of a first class and a sample of a second class contained in a source domain and a sample of the first class contained in a target domain, respectively; a pseudo-sample generation unit configured to generate pseudo-samples of the second class in the target domain based on a distribution of samples of the first class contained in the target domain in a feature space of the features extracted by the feature extraction unit; and a data transformation unit configured to perform data transformation in the feature space by machine learning such that a distribution of samples of the first class and samples of the second class contained in the source domain approximates a distribution of samples of the first class and the pseudo-samples of the second class in the target domain.