EAGER: Rule Induction Games to Explore Differences between Human and Machine Intelligence

这个项目解决了人类和机器学习之间关系中一个以前未被探索的问题。许多挑战人类智力的问题（国际象棋、围棋）已经让位于现代计算机算法。然而，一些对人类甚至动物来说很容易的任务——比如灵活的运动和对周围环境快速而强大的视觉理解——仍然处于人工智能研究的前沿。计算机计算无误。然而，例如，相当多知道奇数和偶数区别的人会说798是奇数，也许是因为它的三分之二的数字是奇数。计算机的学习方式和人类的学习方式有根本的区别吗？我们是否能够通过对游戏的严格研究（游戏邦注：即玩家必须通过试错来学习规则）而找到它们？这个项目使用涉及规则学习的游戏来探索人类和机器学习之间的异同。它将寻求对人类学习的新见解，也可能提高对机器学习的理解。从长远来看，它的目标是更好地整合算法和人类来解决现实世界的问题；当人类和计算机能够相互补充时，它们才能最好地合作。这个项目将寻找对人类来说容易的规则和对人类来说很难的规则之间的一般区别；特别关注的是找到那些难度顺序与机器完全相反的问题。找到这些逆转背后的原理将有助于对问题进行分类。长期目标是混合系统，人类和机器学习集成，以实现诸如医疗诊断，治疗计划等目标。这个项目如果成功，将有助于严格定义为什么一些对人类来说相对容易的学习问题对机器来说却更困难，反之亦然。通过关注规则发现的具体活动，这项研究甚至可能为科学过程提供新的亮点，这一过程被称为“发现自然规则”。该项目采用严格平衡的方法探索机器学习和人类学习之间的互补性，使用“规则归纳”挑战，向人类和计算机提出挑战。计算机将使用最先进的深度神经网络，并探索项目编码语言中可描述的规则的假设空间。心理学研究探讨了跨规则迁移学习、语言和命名在规则发现中的作用等关键问题。人类和机器“玩家”都是通过试错来学习规则的。规则编码语言、强化学习过程和评分系统确保了人类和机器学习者的对称性。绩效衡量标准将包括折扣奖励和趋同于无差错的游戏。学习曲线将用于衡量学习每条规则的难度。实验条件将有系统地变化，不仅包括要学习的规则，还包括显示的最小和最大不同形状的数量，用户在尝试学习给定规则时可能使用的最大“棋盘”数量，以及玩家通过表现获得奖励的激励/奖励结构等参数。该研究将寻求识别对规则的分类，以便对人类来说更容易的类别对计算机来说更困难，反之亦然。该项目将涉及使用多种机器学习方法的广泛实验，以及亚马逊土耳其机器人（Amazon Mechanical Turk）的人类学习表现数据。“比较不同规则的可学习性，可以让我们对人类的学习偏见有新的认识，可能对构建课程很有用，也可能有助于确定哪些知识差距对人类解决问题最有害。”我们的目标是解释或解释这些异常的规则类对与其他规则类的区别，在这些规则类中，人类和计算机的相对难度是相同的。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project tackles a previously unexplored problem in the relationship between human and machine learning. Many problems that challenge human intelligence (chess, Go) have yielded to modern computer algorithms. Yet some tasks that are easy for humans, or even animals — such as flexible locomotion and rapid and robust visual understanding of the surroundings are still at the cutting edge of artificial-intelligence research. Computers calculate without error. Yet, for example, quite a few people who know the difference between odd and even will say that 798 is odd, perhaps because two thirds of its digits are odd. Are there fundamental differences between the way computers learn and way humans learn? Can they be found with a rigorous study of games where the player must learn a rule by trial and error? This project uses games involving the learning of rules to explore similarities and differences between human and machine learning. It will seek new insights into human learning and may improve understanding of machine learning as well. Long-term, it aims to better integrate algorithms and humans for solving real-world problems; humans and computers work together best when they can complement each other, This project will seek generalizable distinctions between rules that are easy for humans and rules that are hard for humans; the special focus is to find problems where the order of difficulty is exactly reversed for machines. Finding the principles behind these reversals will help to triage problems. The long-term goal is hybrid systems, human and machine learning integrated to achieve goals such as medical diagnosis, treatment planning, etc. This project if successful will contribute to rigorously defining how and why some learning problems that seem relatively easier for humans are nonetheless more difficult for machines, and vice versa. With a focus on the specific activity of rule finding, this research may even shed new light on the scientific process, which has been characterized as “discovering the rules of nature.” This project explores complementarity between Machine Learning and Human Learning with a rigorously balanced approach, using a “rule induction” challenge that is presented to both humans and computers. Computers will use state-of-the art deep neural networks, and explore the hypothesis space of rules describable in the project’s coding language. The psychological research investigates crucial problems such as transfer learning across rules, and the role of language and naming in rule discovery. Both human and machine “players” learn the rules by trial and error. The rule encoding language, reinforcement-learning processes, and scoring systems ensure symmetry of human and machine learners. Performance measures will include discounted reward and convergence to error-free play. Learning curves will be used to measure the difficulty of learning each rule. Experimental conditions will be systematically varied, including not only the rule to be learned, but also parameters such as the minimum and maximum number of different shapes displayed, the maximum number of “boards” that a user may use in attempting to learn a given rule, and the incentive/reward structure by which players earn rewards for their performance. The research will seek identify pairs of classes of rules such that the class that is easier for humans is more difficult for computers, and vice versa. The project will involve extensive experiments using diverse machine-learning approaches, as well as Amazon Mechanical Turk for data on human learning performance."Comparing the learnability of different rules sheds new light on human learning biases, may prove useful for structuring curricula, and may help identify which gaps in knowledge are most detrimental to human problem solving. The goal is to interpret or explain what distinguishes these anomalous pairs of rule classes from others where the relative degree of difficulty is the same for humans and computers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.