330 0

Feedback-supervised Learning with A Small Dataset

Feedback-supervised Learning with A Small Dataset
Other Titles
소량 데이터 셋 기반의 피드백 지도학습
Kim, Jin Hui
Alternative Author(s)
Issue Date
The modern society is often called the ‘era of big data.’ The amount of data is growing infinitely due to the development of various systems. And since it is hard for humans to keep track of all those data, machine learning is widely used as a tool for data processing. Machine learning creates models for data analysis, then learn about the data using the generated models, and find patterns within the data. It minimizes human intervention and is suitable for faster data processing. For extensive data analysis, machine learning uses techniques such as ‘classification’ and ‘clustering.’ Classification refers to a form of supervised learning which focuses on classifying the data according to its key characteristics. Clustering refers to a type of unsupervised learning which collects similar data with vague features, then group them into clusters. Recently, deep learning has shown significant improvement in performance in automatic classification. Since deep learning has a high model complexity, it ensures high performance when given a large amount of data. If the amount of initial learning data is not sufficient, the overall performance may drop due to overfitting, which can lead to unwanted results. However, collecting hundreds of thousands of data in the real world consumes a lot of time and is also very expensive. Moreover, even when there is an expert who can guide the criteria for classification, there is no direct way to use human feedback to improve learning performance when the given amount of data is small. This study focuses on improving the classification performance of small amounts of data using the feedback of field experts. Feedbacks from field experts could be made through tags or data augmentation. Such feedbacks are made not only to avoid overfitting due to small data but also to improve the performance through interaction with machine learning by reflecting on human intentions. Also, to prevent the biggest problem of classification when using small data – overfitting -, transfer learning has been used. Transfer learning was performed using the well-known CIFAR-10 data set, and the data set has been applied to the deep learning network classifier before the processing the main data. We have compared the classification performance of two cases: (1) when only used item images, (2) item images used with expert feedback. Experiments have shown that classification accuracy increases when expert feedback is applied to the learning process.
Appears in Collections:
Files in This Item:
There are no files associated with this item.
RIS (EndNote)
XLS (Excel)


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.