Human Perception
Active process of selecting, organizing, and interpreting sensory inputs to construct functional internal representations of the environment. Perception relies on predictive mechanisms and embodied experience rather than passive data recording.
Characteristics
- World Modeling: Humans maintain dynamic world-models encoding physical laws, causal relationships, and spatial geometry, enabling simulation and prediction of future states.
- Sensory Grounding: Perceptual priors are derived from multisensory interaction with the physical environment, anchoring abstract concepts to motor and sensory feedback.
- Active Inference: Perception minimizes prediction error by updating internal models based on sensory evidence; related to Predictive Coding frameworks.
Comparison with AI Systems
Analysis highlights structural divergences between human perception and large-language-models regarding physical reality understanding:
- LLMs primarily operate on statistical token correlations, often lacking grounded semantics or causal understanding of physical dynamics.
- Humans generalize effectively to novel physical scenarios via embodied priors; LLMs exhibit hallucinations and reasoning failures in out-of-distribution physical contexts.
- World Models are investigated as a technical bridge to impart AI systems with simulation capabilities and physical reasoning comparable to human internal modeling World Models: Bridging Human-AI Understanding of Physical Reality.
Related
- world-models
- Embodied Cognition
- Predictive Processing
- Simulator Theory
- Causal Inference