Decision-Making Under Deficiency

Introduction

Color vision is conventionally described as categorical. Objects are "red," "blue," or "green"—discrete labels that the visual system appears to assign with confidence and ease. The rainbow effect in color perception exemplifies what researchers term categorical perception: the wavelength difference between blue and green appears far larger than an equivalent physical difference between two shades within the blue band (Harnad, 1987). This categorical structure seems fundamental to how vision operates.

Yet beneath the surface, color perception—like all perception—is probabilistic. The visual system integrates noisy sensory input with prior experience, producing posterior distributions over possible categories (Knill & Pouget, 2004; Ma & Jazayeri, 2014). For most observers, these distributions peak sharply, creating the phenomenological illusion of certainty. The computational machinery remains hidden precisely because it works so well.

Color vision deficiency changes this picture. When cone photoreceptors are absent, shifted, or functionally reduced, the posterior distributions that typically collapse around single categories instead remain flatter across regions of color space (Bosten, 2019; Isherwood et al., 2020). Judgments become explicit probability statements rather than automatic categorical events. What is usually hidden in typical vision becomes manifest: perception is never absolute, only probabilistic inference—stabilized when confidence thresholds are crossed, unstable when they are not.

This paper proposes that color vision deficiency—reframed here as color differentiation disorder (CDD)—offers a unique window into the computational architecture underlying categorical perception. Rather than treating CDD as perceptual failure, I argue that it reveals the adaptive mechanisms that all perceptual systems employ when operating under uncertainty.

Theoretical Framework

Perception as Probabilistic Inference

The Bayesian framework for perception holds that the brain continuously generates probabilistic estimates of environmental states by integrating sensory evidence with prior knowledge (Knill & Richards, 1996; Kersten et al., 2004). Perception, on this view, is not passive registration but active inference—the brain's current best explanation of the hidden causes of sensory signals (Friston, 2010; Clark, 2013).

In color perception, this inference operates over a continuous stimulus space that must be mapped onto discrete categorical representations. The category adjustment model, developed by Huttenlocher and colleagues (Huttenlocher et al., 1991, 2000), formalizes how observed stimuli are encoded both as fine-grained sensory representations and as category-level prototypes. Reconstruction of the stimulus involves combining these two sources through Bayesian integration, with the relative weighting determined by their respective uncertainties.

Recent extensions demonstrate that category effects in color cognition follow standard principles of probabilistic inference under uncertainty (Regier & Xu, 2017). Cross-linguistic studies confirm that the influence of linguistic categories on color perception scales with perceptual uncertainty: the same stimulus yields stronger categorical effects when presented under degraded conditions (Roberson et al., 2005; Winawer et al., 2007). As fine-grained evidence becomes less reliable, prior categorical knowledge must be weighted more heavily.

Predictive Processing and Precision Weighting

The free energy principle and its instantiation in predictive processing theory provide a complementary framework (Friston, 2005, 2010; Clark, 2013). On this view, the brain maintains a hierarchical generative model of the world and continuously attempts to minimize prediction error—the discrepancy between expected and actual sensory input.

A central concept is precision weighting. Prediction errors are not treated uniformly; they are weighted by their estimated reliability (Feldman & Friston, 2010). High-precision errors are amplified and propagated up the hierarchy. Low-precision errors are attenuated. When sensory signals are degraded, the brain must recalibrate its estimates of precision, becoming more "theory-driven" and less "data-driven"—echoing Helmholtz's description of perception as unconscious inference (von Helmholtz, 1867/1962).

Color Vision Deficiency as Model System

Color vision deficiencies provide an ideal natural experiment for investigating perception under chronic uncertainty (Isherwood et al., 2020). In anomalous trichromacy, one class of cone photoreceptors has an altered spectral sensitivity curve that reduces the effective separation between channels (Neitz & Neitz, 2011). The result is a permanent reduction in the signal-to-noise ratio for chromatic information along specific axes of color space.

Recent neuroimaging research shows substantial plasticity in response to CDD. As predicted by the receptor-level deficit, primary visual cortex displays reduced chromatic responses in anomalous trichromats. However, later visual areas show responses closer to those of typical observers (Tregillus et al., 2021), consistent with post-receptoral amplification—a compensatory gain adjustment that partially restores the weakened chromatic signal.

The Dual-Process Architecture

Categorical perception appears to operate through two functionally distinct processing modes—what I term phenomenological predictors and actuarial predictors. These modes are not sequential stages but parallel resources that the system deploys differentially depending on its current level of confidence.

Phenomenological Predictors

Phenomenological predictors characterize the system's operation in a state of active uncertainty. When confidence is low—when the posterior distribution over categories remains relatively flat—the perceptual system continues sampling, maintaining multiple candidate interpretations simultaneously. This is not an error state but an adaptive response: the system is "staying open" rather than committing prematurely to a potentially incorrect categorization.

The phenomenology of this mode is characterized by instability, provisionality, and effortful engagement. In color perception for individuals with CDD, certain regions of color space may chronically trigger phenomenological processing: the stimulus is perceived, but its categorical identity remains uncertain and subject to revision.

Actuarial Predictors

Actuarial predictors describe the system's operation once sufficient evidence has accumulated for commitment. The term actuarial is borrowed from risk-assessment contexts in which probabilistic reasoning is formalized and validated (Meehl, 1954). In perception, actuarial predictors integrate contextual and structural cues—lighting conditions, surface properties, surrounding colors, object identity—to stabilize categorical assignment.

When actuarial predictors dominate, the perceptual system operates efficiently. Prediction errors are minimized; the generative model's current explanation is adequate. The system no longer needs to engage in costly active sampling because confidence has surpassed the threshold required for commitment.

The Compilation Mechanism

The transition from phenomenological to actuarial processing is mediated by what I term compilation—the process through which provisional categorical representations consolidate into stable, low-cost knowledge structures. Compilation occurs when confidence crosses a threshold: enough evidence has accumulated for the system to commit without ongoing active sampling.

The metaphor comes from computer science, where compilation refers to transforming high-level code into efficient machine instructions (Aho et al., 2006). In perception, compilation similarly transforms flexible, resource-intensive inference into efficient, automatic categorization. A compiled category no longer requires continuous verification; it can be deployed rapidly and with little cost.

Extensions: Anxiety, Credibility, and Trust

The dual-process architecture developed for color perception extends naturally to other domains where categorical judgment must emerge from uncertain evidence.

Anxiety as Phenomenological Processing

Anxiety can be reconceptualized as the phenomenological experience of unresolved uncertainty. When the predictive model cannot adequately explain incoming signals—when prediction errors persist and confidence thresholds are not crossed—the system remains in an effortful, resource-intensive state. The subjective experience of this state is anxiety: the sense that something remains unresolved and requires continued vigilance.

On this account, anxiety is not inherently pathological. It is the phenomenological signature of adaptive uncertainty management—the system's refusal to compile prematurely when the evidential base is insufficient. What becomes pathological is the failure of compilation: a chronic inability to cross confidence thresholds even when conditions would support doing so.

Credibility as Compilation

Credibility assessment presents the same structural challenge: determining when a source is trustworthy given incomplete and noisy evidence. Phenomenological mode corresponds to active evaluation—weighing signals, maintaining alternative interpretations, delaying commitment. Actuarial mode corresponds to compiled credibility judgments—trust or distrust becomes automatic, reducing computational load.

Trust as Resource Allocation

Trust describes the willingness to allocate resources based on compiled credibility judgments. When trust is high, prediction errors from the trusted source are granted high precision weighting and integrated with minimal verification. When trust is low, the system must keep evaluation active, preserving flexibility at the expense of computational cost.

Phenomenological Data: First-Person Report

The Phenomenology of Instability

I describe my color deficiency as full-spectrum. Difficulty distinguishing colors increases as hue moves away from primaries and secondaries; instability is greatest near the 50% gradient—the midpoints where categories blur. These are the regions where probability distributions flatten, where the visual system cannot assign confident labels, where phenomenological processing dominates.

What does instability actually feel like? It feels like nothing definite. The experience is not of seeing a wrong color or a missing color, but of seeing a color that refuses to resolve. The percept is provisional, contingent, revisable. It is the difference between a statement and a question—between "this is green" and "is this green?" The question mark is not silent; it registers phenomenologically as a kind of openness, an incompleteness that does not close.

Context as Inference Engine

I do not think in terms of color. Almost never. Not in search functions, not in spontaneous categorization, not in the automatic labeling that I understand typical observers perform constantly and without effort.

When I look at an apple, I do not think "that's really red." I think "that's an apple." The object identity comes first. The color follows as inference. The apple is red because apples are red—the category supplies the color, not the percept. This inversion represents a fundamental reallocation of the inference process: object recognition runs the classification, and color becomes a downstream consequence rather than a primary input.

The Gradient Problem

Consider a paint gradient running from blue to green. For a typical observer, there is a point—perhaps abrupt, perhaps gradual—where blue becomes green. The boundary exists. It may be fuzzy, but it is locatable.

For me, that boundary is everywhere and nowhere. The gradient does not transition; it destabilizes. The region of uncertainty is not a line but a zone—a broad band where categories compete without resolution. Within that zone, the system samples continuously. It does not commit. The posterior distribution is flat; confidence remains below threshold; the category does not compile.

Conclusion

This framework reconceptualizes deficiency not as perceptual failure but as a window into the computational architecture that underlies all categorical judgment. The same architecture appears across anxiety, credibility assessment, and trust formation, suggesting a general principle:

Systems adapt by reasoning through instability, building structure from noise, conserving resources once confidence is secure, and allocating computational effort where uncertainty persists. The question is not whether instability exists—it exists in all perception—but whether the system can navigate it adaptively: staying open when evidence is insufficient, committing when it is adequate, and knowing the difference.