Speaker: Prof Suzuki Takayuki, The University of Tokyo
Abstract:
Classical AI, which has been dominant until 1980s in AI research, regarded thinking as computation, that is, formal manipulation of symbols. Though this approach worked well in simple tasks, it has repeatedly failed in dealing with difficult tasks. One fundamental problem is that it is extremely hard for programmers to identify rules and features necessary for difficult tasks. With the use of machine learning and neural networks, contemporary AI has overcome this difficulty. It is not clear, however, we can build human-like artificial intelligence with this approach. The reason is that the approach of contemporary AI is effective only when a task is not open-ended and large training data are available. Whether and how we can build human-like artificial intelligence without the process of massive trial and error remains to be an important challenge to contemporary AI.
