Core Concepts

The human mind represents the direction of gravity as a Gaussian distribution centered around the vertical direction, rather than as a singular downward vector. This stochastic world model on gravity provides a better explanation for stability inference and the illusion that taller objects are perceived as more unstable.

Abstract

The study investigates how the physical law of gravity is embodied in the brain as a world model that guides inferences on objects' stability. The key findings are:
The world model on gravity is not a faithful replica of the physical world, but rather a stochastic model that captures the essence of the vertically downward direction of gravity as the maximum likelihood of a Gaussian distribution.
The stochastic feature of the world model not only fits humans' stability inference behavior better than a deterministic model, but also provides new insight into the daily illusion that taller objects are perceived as more likely to collapse.
A reinforcement learning framework illustrates how the stochastic feature of the world model on gravity can emerge through interactions with the physical environment, without the need for external noise or perturbations.
The stochastic world model balances accuracy and speed in stability inference, providing an ecological advantage for survival in the physical world. Humans are found to perform a limited number of mental simulations before making stability judgments.
In summary, the stochastic world model on gravity provides an example of how a priori knowledge of the physical world is implemented in the human mind, enabling adaptive functionality in open-ended environments.

Stats

The number of possible configurations for a stack of 10 blocks is at least 1.14×10^50.
The variance of the Gaussian distribution representing gravity's direction in the world model ranges from 11.1 to 37.1 across participants.
The inference bias, or the difference in stability estimates between participants and the natural gravity simulator, is -0.31.
The inference bias between the mental gravity simulator and the natural gravity simulator increases monotonically with the height of the stack.

Quotes

"The world model on gravity is unlikely to be a faithful replica of the physical world; instead, it encodes gravity's direction as a Gaussian distribution with the vertical direction as the maximum likelihood."
"The stochastic world model on gravity provides a more concise explanation for the daily illusion that taller objects are perceived as more likely to collapse, without assuming external perturbations."
"The world model representing gravity's direction in a Gaussian distribution can emerge automatically as the agent interacts with the environment, without the need for any external perturbation."

Key Insights Distilled From

by Huang,T., Li... at **www.biorxiv.org** 12-31-2022

Deeper Inquiries

The stochastic world model on gravity could be implemented at the neural level in the brain through a combination of neural networks and probabilistic models. Neurons in the brain could encode the different components of gravity's direction (such as vertical and horizontal components) and their probabilities. This information could be processed in neural circuits that integrate sensory inputs, internal simulations, and prior knowledge to generate predictions about the stability of objects. The stochastic nature of the model could be represented by neural populations that exhibit variability in their firing rates or patterns, reflecting the uncertainty in gravity's direction. Additionally, neural mechanisms such as synaptic plasticity and reinforcement learning could be involved in fine-tuning the representation of gravity's direction based on feedback from the environment.

Other physical laws or intuitions that might also be represented in the mind as stochastic models include principles related to friction, elasticity, and fluid dynamics. For example, the perception of friction between surfaces could be represented as a Gaussian distribution, with variations in the coefficient of friction influencing stability judgments. Similarly, the elasticity of materials could be encoded probabilistically to account for uncertainties in how objects deform or rebound upon impact. In the case of fluid dynamics, the behavior of fluids in containers or flowing through channels could be modeled stochastically to capture the inherent variability and unpredictability of fluid motion.

The principles behind the stochastic world model on gravity could indeed inform the development of more flexible and adaptive artificial intelligence systems. By incorporating stochastic representations of physical laws into AI models, such systems could better handle uncertainty and variability in real-world environments. For example, in robotics applications, AI systems could use stochastic models to make more robust predictions about the stability of objects or the effects of external forces. This could lead to AI systems that are better equipped to adapt to changing conditions, learn from experience, and make decisions in complex and dynamic environments. Additionally, the ecological advantage of balancing accuracy and speed demonstrated by the stochastic world model could inspire the design of AI algorithms that prioritize efficiency while maintaining high performance.

0