Cost-Effective Active Learning for Relational Few-shot Learning

• Main Authors: Chang-Sin Dai, 戴長昕
• Other Authors: 廖世偉
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/ka5cn7

碩士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 106 === In recent years, deep learning is developed in tremendous speed nowsdays. Giving the credits to the increase in computing power, a large number of researchers have invested in research, making deep learning shine in many fields such as machine vision, natural language processing, and speech processing. While meta-learning and few-shot learning have not yet been effectively applied in industry, their ideas and creativity are at the forefront of deep learning, and its development is still attracting attention. In this thesis, We found that active learning and few-shot learning have very similar ideas. Based on the concept of active learning, we design a suitable training sample selector for small sample learning tasks to optimize model performance.we use a relational network model that simulates the human recognition process, combined with active-learning-like method to reduce training costs and make it efficient, and let the model learn better meta knowledge to deal with unknown tasks. We design experiments to explore the effects of various parts of the method. By analyzing the experimental results, we hope to be able to open the black box in the deep structure model with less sample learning.