eprintid: 66072 rev_number: 11 eprint_status: archive userid: 2047 dir: disk0/00/06/60/72 datestamp: 2021-02-05 10:31:49 lastmod: 2021-02-05 10:31:49 status_changed: 2021-02-05 10:31:49 type: thesis metadata_visibility: show creators_name: Bang, An contributors_name: Wu, Wenjun contributors_name: Ferré Grau, Xavier title: Research on interactive machine learning method based on group intelligence competition rights: by-nc-nd ispublished: unpub subjects: informatica full_text_status: public keywords: Group intelligence, Machine learning, Active learning, Artificial intelligence abstract: As time goes by, the current field of artificial intelligence is undergoing profound changes. In the past, we only focus on experts but now we start to pay attention to group intelligence. Therefore, the theory and method of group intelligence will gradually become one of the core areas of a new generation of artificial intelligence, meanwhile it has a basic and supportive role in other research fields of artificial intelligence. Based on group intelligence, the machine learning models emphasize the concept of interactive machine learning, which means placing human in computing loops. On the one hand, human can select data sample heuristically through their subjective recognition abilities, on the other hand, computers can iterate algorithms fast through machine analysis. By this way, the models of machine learning can reach human expected results in a shorter time with less data. This paper innovatively proposes an interactive machine learning method based on the concept of group intelligence, which combines traditional active learning with deep learning algorithms called Deep Active Learning. It is proved by experiments that the method can achieve the higher training precision that traditional machine learning model can't achieve by using around 50% of the training dataset compared with traditional methods. It solves the problems that traditional active learning needs to select feature extraction algorithm and traditional deep learning depends on the scale of training dataset deeply. In the end, the research results have been applied successfully, it is combined with the group intelligence competition, finally forms the online platform called BUAA Data Lab. The platform provides reliable support for machine learning competitions, machine learning model optimization, machine learning algorithm experience exchange and data sharing in campus. Now the platform has been successfully applied to teaching in Beihang Universtity, and has reached cooperation with the national “The Share Cup” science and technology resource sharing service innovation competition. date_type: completed date: 2019-05 pages: 83 institution: ETSI_Informatica department: Lenguajes thesis_type: masters master_title: Ingeniería del Software citation: Bang, An (2019). Research on interactive machine learning method based on group intelligence competition. Thesis (Master thesis), E.T.S. de Ingenieros Informáticos (UPM) . document_url: https://oa.upm.es/66072/1/TFM_AN_BANG.pdf