Intuitive Physics: Current Research and Controversies

Summary

• Early research in intuitve physics indicated that humans exhibit common misconceptions and biases when predicting the physical world
• However, more recent work has indicated that some biases can be explained by the application of normative physical principles to noisy perceptual inputs
• How these physical principles are learned, represented, and applied remains unclear
• Links: [ website ] [ pdf ]

Background

• Early work found that human predictions often disagree with (ground-truth) Newtonian physics
  • Participants asked to draw how situation would unfold based on static diagram of physical scenario
• These misconceptions can be reduced by changing the experimental paradigm
  • Adults are bad at drawing trajectory, but can predict landing position
  • Suggests that humans have strong intuitive “physics engine”, but only applied in certain conditions
  • Cortical activation associated with explicit physical knowledge does not enitrely overlap with areas for tacit physical inference

Methods

• Recent research on intuitive physics uses the noisy Newton framework, where inference is achieved by passing noisy information through a physics engine

  • Object dynamics knowledge is “written in”
  • The belief that humans contrust mental models for physical situations underlies these models

• Shift metric to correlation of predictions between humans and models versus absolute performance levels

• Account for human predictions through noisy perceptual inputs, since physical princeiples are approximated but not biased

• Additional hybrid approach combining knowledge-based physics model with learned network for predicting physical attributes from visual inputs

• Formal definition of problem

• Model details/experimental setup

Results

• Probabilistic simulation model predictions correlate well with human performance across experimental conditions for stacked blocks and liquids moving past obstacles
  • Introducing randomness into dynamics improves fit to human predictions of an occluded object bouncing in a box
• Support for core knowledge thesis: core physical principles guide the construction of tacit theories of motion
  • Initial knowledge about the physical world is specific to learned domains

Conclusion

• Intuitive physics research has benefited from advances in:
  • Stimulus displays - from static diagrams to dynamic animations
  • Computational theory - from heuristic accounts to probabilistic simulation framework
  • Choice of physical situations - from focus on rigid objects to non-rigid fluids
• While human predictions are consistent with probabilistic inference, such models require numerous simulations based on hard-coded physical principles