COMPCOG: Intuitive Physics without Intuition or Physics: Leveraging Deep Neural Networks to Model Human Physical Reasoning

拟议研究的主要目的是利用人工智能（AI）和深度学习的最新进展，深入了解人类认知的内部运作。虽然人类感知和认知的本质只能从行为中推断出来，但模仿人类行为的深度神经网络提供了一个独特的机会，可以仔细检查系统的内部运作，并有可能理解系统是如何以及为什么以这种方式工作的（以及人类系统是如何以及为什么以这种方式工作的）。虽然认知科学一直使用正式的计算模型，但人工智能的认知能力第一次开始在几个领域与人类竞争或超过人类，这为使用这些模型正式理解复杂的人类行为提供了机会。具体来说，在拟议的研究中，研究人员将使用深度神经网络模型来更好地理解直观的物理推理——我们理解环境中物体行为的能力。例如，人们一眼就能判断出一堆盘子是否即将落下，或者它是否稳定。这种类型的直觉物理推理是人类认知的一个基本方面，它是我们对世界上物体的日常推理的基础，但也可能支持物理科学的正规教育（例如，学习牛顿物理学）。因此，拟议研究的总体目的是利用深度学习的最新进展来发展对人类直觉物理推理的丰富基础理解，这最终有可能影响正式的STEM教育。在提议的工作中，研究人员将结合认知科学（理想观察者分析）和人工智能（深度卷积神经网络）的工具，来测试一种新的心理学理论（直观物理推理作为形状空间的外推），以深入了解人类认知的一个核心方面（直观物理推理）。这项工作包括测试人类参与者，以获得人类在各种直观物理判断上的表现基准，然后训练深度神经网络模型来执行相同的任务，在某些情况下，被限制为理想观察者（执行最佳计算来执行任务的系统）。通过研究人类与这些深度神经网络模型之间的异同，并仔细研究以人类水平的精度执行任务的模型的内部工作原理，研究人员将开发基于人类直觉物理推理的神经网络模型。因此，这项工作提出了认知科学和人工智能之间的跨学科整合，有可能影响基本认知理论及其在正式STEM教育中的应用，并增加机器视觉、人工智能和旨在模仿人类智能的自主代理领域之间的协同作用。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。

The broad purpose of the proposed research is to leverage recent advances in artificial intelligence (AI) and deep learning to gain insight into the inner workings of human cognition. While the nature of human perception and cognition can only be inferred from behavior, deep neural networks that emulate human behavior offer a unique opportunity to scrutinize the inner workings of the system and potentially understand how and why the system works the way it does (and, thus how and why the human system works the way it does). While cognitive science has always employed formal computational models, for the first time the cognitive abilities of artificial agents has begun to rival or exceed those of humans in several domains, providing the opportunity to formally understand complex human behavior using these models. Specifically, in the proposed research the investigators will use deep neural network models to better understand intuitive physical reasoning — our ability to understand the behavior of objects in our environment. For example, at a glance, people can judge whether a stack of plates is about to fall, or whether it’s stable. This type of intuitive physical reasoning is a fundamental aspect of human cognition which underlies our everyday reasoning about objects in the world, but likely also supports formal education in the physical sciences (e.g., learning Newtonian physics). Thus, the general purpose of the proposed research is to leverage recent advances in deep learning to develop a rich basic level understanding of human intuitive physical reasoning, which ultimately has the potential to impact formal STEM education.In the proposed work, the investigators will combine tools from cognitive science (ideal observer analyses), with tools from artificial intelligence (deep, convolutional neural networks), to test a novel psychological theory (intuitive physical reasoning as extrapolation in shape space) to gain insight into a core aspect of human cognition (intuitive physical reasoning). This work involves testing human participants to obtain human performance benchmarks on a variety of intuitive physical judgments, then training deep neural network models to perform the same tasks, in some cases constrained to behave as ideal observers (systems that perform the optimal computation to perform the task). By examining similarities and differences between humans and these deep neural network models, and scrutinizing the inner workings of models that perform the task with human-level accuracy, the investigators will develop neural network based models of human intuitive physical reasoning. As such, this work presents an interdisciplinary integration between cognitive science and artificial intelligence, with the potential to impact basic cognitive theory and its applications to formal STEM education, and to increase the synergy between the fields of machine vision, artificial intelligence, and autonomous agents that aim to emulate human intelligence.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.