:::

Explainable Artificial Intelligence

 1. Coordinators and core members 

  Jason Chang (
張俊盛 Computer Science, NTHU) 
jason@nlplab.cc
  Daw-Wei Wang (王道維 Physics, NTHU)  dwwang@phys.nthu.edu.tw

 II. Core Members:

       Von -Wunun Soo (蘇豐文 Computer Sciences, NTHU)
        Po-Chung Chen (陳柏中 Physics, NTHU)
        Che-Rong Li (李哲榮 Computer Science, NTHU)
        Ted Kuo (郭志義 Computer Science, NCTU)
        Yu-Chen Chan (詹雨臻 Learning Science and Technology, NTHU)
        Wei-Tian Tsai (蔡維天 Linguistics, NTHU)
        Su-Yen Chen (陳素燕 Learning Science and Technology, NTHU)
        Shan-Ru Lin (林珊如 Education, NCTU)

III. Major Direction: 

       It is known that recent rapid progress of AI application are mainly promoted by the success of Deep Artificial Neuron Network (DANN) or Deep Learning (DL), which can be trained to find important hidden features inside the Big Data created (and sometimes annotated) by human. However, even though that, in many applications, shows great precision in prediction and classification, the results provided by ANN or DL are almost unexplainable due to the huge number of fitting parameter inside. In other words, the effectiveness of these AI application will be limited by the machine’s current inability to explain their decisions to human users. In fact, DARPA of USA has called for proposals for XAI research in August 2016, showing the importance of this particular aspect in AI application.

        From physics point of view, such explanation can be understood as the process for extracting an effective model from a complex system, identifying the most important key factors after certain renormalization. Therefore, it can be a very proper and important interdisciplinary direction for researchers in theoretical physics, computer science, education, and other humanity disciplines. So far there is no such research proposal in Taiwan and the IDP program supported by NCTS will be a leading player for the whole community.

        In this XAI program, we shall focus on the following areas:
 
        (1) Features and Logic Identification, as a new machine learning process:
        The present DL algorithm relies on a simulation through thousands or more parameter tuning for high classification effectiveness. However, probably there are much redundancy and only certain key combinatorial features are important for true application. We will then develop some algorithm to capture these key features and transform them into understandable model for a logic relationship. This can be done from some simpler case with known theory. Physical models and related information theory can be attempted for this purpose.

        (2) Natural Language, as a model of explanation:
        Since any “explanation” shall be formulated to human language for the communication purpose, understanding the structures and features of spoken or written language is an important step. There has been many important contribution and achievement in linguistics for the theory of common language, and therefore it can be combined with current algorithm for a model of explanation. In fact, there could be several levels of understanding and each of them implies different mental processes.

        (3) Expression and Education, as an application interface:
        One of the most important application of XAI can be in education, where students or users can be educated by AI through explainable results. This could greatly enhance the development of next-generation learning technology. How to develop a Human-Computer-Interface (HCI) to enhance the communication effects of a explainable AI can be the important factor of social impact.

IV. Tentative plan of activities

      We will host activities fostering inter-disciplinary collaboration under the theme of XAI. They include tentatively the following events.

      (1) Regular seminar talks for speakers from various research institutions.
      (2) One or two mini-workshops per year
      (3) Hosting short-term local or international visitors. The visitors may coincide with the speakers of the conference.
      (4) Travel expenses for postdocs or students to international or domestic conferences.

cron web_use_log