Artificial Intelligence or Natural Stupidity? Deep Learning or Superficial Teaching?

摘要

Go is an ancient board game in which two players, by placing "stones" on a square grid, aim to surround more territory than the opponent. It was a pivotal moment in the history of humankind when AlphaGo Master, a computer program developed by DeepMind Technologies from the United Kingdom, defeated professional Go player Ke Jie in three games of Go during the 2017 Future of Go Summit in Wuzhen, China. At the time of the match, Ke Jie was the number one player as ranked by the Chinese Weiqi Association, the Japan Go Association, and the Korea Baduk Association, and was considered to be the best Go player in the world. Perhaps less dramatic, but equally important, are the moments when Boston Dynamics' bipedal humanoid robot named Atlas, picked up a box from the ground and then moved it onto a shelf or when another humanoid robot, Honda's Asimo (for "Advanced Step in Innovative MObility") walked to a table, poured juice from a thermos into a cup, and then served the drink. These apparently ordinary tasks, which utilize only low-level sensorimotor skills, actually require more computational resources than do high-level reasoning tasks like playing chess. This is an observation made decades ago by researchers working on artificial intelligence (AI) and robotics, and is termed "Moravec's paradox" (Moravec, 1990). Robots with artificial intelligence and sensorimotor skills are surpassing humans in the performance of many tasks and doing so at a surprising pace. Many analyses have thus been done to identify the human jobs that are at risk of being replaced by robots (Frey and Osborne, 2013; Chui et al., 2016; Berriman and Hawksworth, 2017; Makridakis, 2017). At least according to one analysis, jobs in the wholesale and retail trades are at the highest risk, while those in education are at the lowest (Berriman and Hawksworth, 2017). It appears that we anatomy educators are safe for now, or are we? Machines that can teach have been around for almost a hundred years, but it was not until the recent advent of deep learning models based on artificial neural networks that these machines can be called intelligent. The teaching machine that Sidney Pressey of The Ohio State University designed in 1924 looked like a typewriter (Fry, 1960). It showed a question through a window to the student, who then answered the question by pressing a key corresponding to the answer that the student chose. Through a mechanical process, the machine recognized whether the answer was correct. If it was, the machine then moved to a new question. The machine could not be considered intelligent since it could only "recognize" one kind of data input (the pressing of a key) and then "respond" mechanically. The new "intelligent" machines are very different. They can take in complex information like human speech, static and dynamic images, or financial data and transform this information into more abstract and composite responses. They are called "deep" learning systems because of the many layers of transformation between input and output, and have diverse applications in speech recognition, computer vision, medical image analysis, autonomous driving, board games, financial fraud detection, etc. Given this level of development, can intelligent machines teach anatomy? Can they replace anatomy educators? Until quite recently, these questions would seem nonsensical, since no attempt had been made to produce such machines. Most intelligent tutoring systems (ITSs) are in disciplines where content materials are digital or can be easily digitized, for example, mathematics (Beal et al., 1998; Craig et al., 2013), computer programming (Mitrovic, 2003), engineering (Zakharov et al., 2005), reading (Heffernan et al., 2006), or music (Miletto et al., 2005). Anatomy, however, deals with the internal structure of organisms. But even that has been digitized to a high degree of accuracy (Ackerman, 1998; Park et al., 2005; Tang et al., 2010). Significant pedagogical innovations have bee