In this project, we will explore whether children and adults can learn non-intuitive physicalforces, particularly those that violate our core knowledge about the physical world, and whether thislearning varies as a function of age. For example, can children and adults who are exploring andinteracting with VR worlds learn that pushing an object will actually draw the object closer rather thanfurther away? Further, will children, who have less experience interacting with objects, learn thesenon-intuitive forces quicker or more slowly than adults? This interdisciplinary research proposal drawsfrom developmental psychology, cognitive science, and AI. Data from this project will not only advanceour understanding of learning, specifically in the possible differences in children and adults’ learning, butcan also be used to advance our understanding of AI. In particular, we will test whether agents withvarious learning mechanisms show similar patterns to humans in learning intuitive vs counterintuitivephysics in the same virtual environments. In addition, the data from this project can be used as trainingvideos for robotic tasks, and also speaks to practical questions about how children and adults may learnthrough VR environments.
Overview
Novelty and Innovation: The consistent physics of the real world means that a learned model will alwaysappear to be similar to one that evolved. This makes it difficult to resolve debates in AI about the relativeimportance of prior core knowledge that is ‘built in’ vs. learning during development. A crucial test wouldbe to explore whether agents in an alternative world with different physics would be constrained by anevolved core knowledge or would learn the counter-intuitive system. Advances in VR mean that, for thefirst time, these experiments are possible. Specifically we will test whether humans can learncounter-intuitive principles of physical causality, object permanence and gravity. Using VR will allow us toexplore how children and adults can adapt to radically novel physical environments.
Technical Objective: Our data will have various metrics as output from the VR situational games. As inDubey et al., this data can inform computer scientists about the crucial core systems or learningmechanisms that need to be baked into AI systems in order for them to efficiently learn in a way that issimilar to children’s learning. Our proposed timeline projects that we will start testing in children in Spring2021. We have received full IRB approval, and have found designated students with VR physics engineexperience to build the actual worlds.
Potential for collaboration: This project is in collaboration with Jitendra Malik and James Hillis. Jitendrais Research Director at FAIR-Menlo Park and a professor of computer science at UC Berkeley,specializing in computer vision. James Hillis is a perception scientist at Oculus Research. Their uniqueresearch areas make them the perfect collaborators at Facebook to pursue our interdisciplinary researchquestions. Alison Gopnik is a professor of child development in psychology with a joint appointment inBAIR, two of her PhD students will be the main researchers on this project Eliza Kosoy and Katie Kimura.Jasmine Collins is a PhD student under Jitendra with a background in computer science andneuroscience. The interdisciplinary nature of the team will allow us to distribute the results ininterdisciplinary research communities. We plan to distribute the simulations we develop so people allover the world can gather more data.
Researchers
- Eliza Kosoy; UC Berkeley: (eko@berkeley.edu);
- Jasmine Collins; UC Berkeley: (jazzie@berkeley.edu),
- Katie Kimura; UC Berkeley: (katieckimura@gmail.com),
- Alison Gopnik; UC Berkeley: (gopnik@berkeley.edu)
- Jitendra Malik: FAIR (malik@eecs.berkeley.edu),
- James Hillis; Oculus (james.hillis@oculus.com)
Links
Contact email: Eko@berkeley.edu
This project was not completed due to Covid-19, all in-person resarch with human children was suspended and we were not able to complete this research.