Innateness and Bayesian Visual Perception

Underdetermination and the Inverse Problem

  • Visual perception reconstructs three-dimensional scene structure from two-dimensional retinal input.

  • Perception faces an underdetermined "inverse problem": infinitely many 3D scenes can project the exact same 2D retinal image.

  • Key ambiguities include depth/size trade-offs and lightness reflectance vs. illuminance (where retinal light intensity represents the product of illumination and reflectance).


Examples of visual underdetermination in depth geometry and surface reflectance

Unconscious Inference and Coincidence Avoidance

  • Helmholtz's Principle of Unconscious Inference: Perceptual systems infer objects and events most likely to produce the received sensory input under normal conditions.

  • Coincidence Avoidance: Visual mechanisms discount interpretations that rely on accidental viewpoints or accidental alignments (e.g., generic viewpoint assumptions, modal and amodal completion).

Bayesian Framework for Visual Perception

  • Visual perception models optimal reasoning under uncertainty using Bayes' theorem:   p(Scene∣Image)=p(Image∣Scene) p(Scene)p(Image)p(\text{Scene} \mid \text{Image}) = \frac{p(\text{Image} \mid \text{Scene}) \, p(\text{Scene})}{p(\text{Image})}

  • Likelihood Function p(Image∣Scene)p(\text{Image} \mid \text{Scene}): Captures physical optics and projection parameters.

  • Prior Distribution p(Scene)p(\text{Scene}): Encodes innate or learned natural constraints about the structure of the physical environment.

  • Overhead Illumination Prior: Assumes light originates from overhead, leading top-to-bottom lightness gradients to be perceived as convex bumps or concave dents.


Egg carton visual array showing convex and concave shape perception driven by overhead illumination assumptions

  • Hershberger (1970) demonstrated that chickens raised with bottom-up lighting still default to an overhead light assumption, confirming its innate origin.

Reconciling Nativism and Development

  • Tuning Bayesian Priors:

    • Innate knowledge operates as default variable distributions or factory settings (p(Hypothesis)p(\text{Hypothesis})).

    • Interaction with environmental statistical regularities continuously updates, tunes, and re-weights prior distributions without discarding initial innate structures.

  • Spontaneous Endogenous Activity:

    • Early neural circuitry (such as ocular dominance columns and orientation selectivity) is structured before sensory exposure via spontaneous, internally generated signals.

    • Postnatal sensory experience acts upon these exact same neural mechanisms to refine visual tuning, demonstrating that innateness and developmental plasticity share identical underlying processing hardware.