Your Brain's Rendering Pipeline,
in 886 Lines of GLSL

GLSL (OpenGL Shading Language) is a C-like language that runs directly on the GPU. A fragment shader is a program that executes once per pixel, in parallel, every frame — making it ideal for real-time image processing. The entire foveated vision simulation runs as a single fragment shader at 60fps.

Vision is an information processing pipeline. At each stage, the signal is transformed: retinal ganglion cells decompose spatial frequencies. The LGN gates what reaches cortex. V1 pools features across receptive fields that grow with eccentricity. V4 reshapes color and form. The question that drives this project — the same question behind any good recommender system or search engine — is: what happens to the signal at each stage? What's preserved, what's lost, and what does the representation look like when it arrives downstream?

Scrutinizer makes those transformations explicit. Each biological stage maps to a signal-processing operation in the shader: Difference-of-Gaussians for ganglion cell decomposition, per-band M-scaling for cortical magnification, saliency-weighted suppression for the LGN gate. The data flows through, and at each stage you can see exactly what it looks like.

Retina → LGN → V1 → V4 · implemented in a single GLSL fragment shader

Signal transformation pipeline: raw photoreceptor input → LGN filtering gate → V1 feature extraction → reassembled perceptual output
The signal at each stage. Dense photoreceptor input is filtered by the LGN gate, decomposed into oriented features by V1, and reassembled into a sparse perceptual representation. Information is progressively refined — not lost, transformed.
Scrutinizer browser demo — foveated vision simulation in real time

The simulation running live on a browser page. Follow the cursor — that's your fovea.

Wrapped in a browser built with the popular Electron framework, the shader runs on any web page in real time — turning the screen into a window onto your own visual processing.

The Pipeline: Biology → GLSL

The visual system is vastly more complex than any current simulation captures — dozens of cortical areas, parallel processing streams, recurrent feedback loops. The shader models four stages chosen for their outsized impact on peripheral appearance: retinal spatial filtering, LGN attentional gating, V1 feature integration, and V4 color processing. Eccentricity from the gaze point drives every parameter.

Retina
Photoreceptors → Ganglion cells. 1:1 foveal, 100:1 peripheral.
sampleSource(uv)
MIP pyramid generation
LGN
Attentional gating. Structure mask. Saliency gate.
V1
Feature extraction. Crowding. Positional uncertainty.
V4
Color processing. Rod vision. Aesthetic rendering.
The Central Equation: Eccentricity

Every stage is driven by one value — the normalized distance from the gaze point:

$$\text{eccentricity} = \frac{|\vec{P}_{\text{fragment}} - \vec{P}_{\text{gaze}}|}{r_{\text{fovea}}}$$

$\vec{P}_{\text{fragment}}$ is the screen position of the current pixel (in GPU terminology, a fragment — the shader runs this equation independently for every pixel on screen, every frame). $\vec{P}_{\text{gaze}}$ is the current gaze point (mouse position or eye-tracker input). At eccentricity = 0, you're at the fovea (full detail). At 1.0, you're at the foveal boundary. At 2.5, parafovea. Beyond that: peripheral vision, where most of the visual field lives — and where most of the shader's work happens.

Biological Primer

Much of what we know about these structures has been understood for decades — Kuffler described center-surround receptive fields in 1953, Hubel and Wiesel mapped orientation selectivity in 1962. The sections below summarize the minimum context for following the simulation. For deeper treatment, see Wandell's Foundations of Vision (1995), or this project's annotated bibliography, which maps each cited paper to its role in the shader implementation.

The Retina: Where Vision Begins

The retina is not a passive sensor — it's a layered neural circuit. Light hits photoreceptors (rods and cones), which convert photons to electrical signals. These signals pass through bipolar cells to retinal ganglion cells, whose axons form the optic nerve. In the fovea — a 1.5mm pit at the retina's center, densely packed with cones — each cone connects to its own ganglion cell (1:1 wiring). In the periphery, 100+ rods converge onto a single ganglion cell, trading spatial resolution for sensitivity. This convergence ratio is the fundamental reason peripheral vision is blurry.

Receptive Fields and Center-Surround

Kuffler's 1953 experiment was elegantly simple: he recorded from individual ganglion cells in the cat retina while shining small spots of light at different positions. He found that each cell responded to a specific patch of the visual field — its receptive field — and that a spot in the center excited the cell while a spot in the surround inhibited it (or vice versa). This center-surround antagonism is the retina's first act of computation: extracting contrast, not absolute brightness.

A neuron's receptive field is the region of visual space that drives its response. Retinal ganglion cells have center-surround receptive fields (Kuffler, 1953): an excitatory center ringed by an inhibitory surround, or vice versa. This makes them bandpass filters — they respond to spatial frequencies matching their field size, not to uniform illumination. The Difference-of-Gaussians (DoG) is a mathematical model of this response profile. Critically, receptive field size scales with eccentricity: larger fields in the periphery means coarser frequency tuning further from fixation.

Receptive Field Mosaic

Each circle represents one ganglion cell's receptive field. Dense, small fields at the fovea; sparse, large fields in the periphery. This is why peripheral vision is blurry: fewer cells, each pooling a wider region.

Photoreceptor Convergence

Photoreceptors (top) funnel onto ganglion cells (bottom). At the fovea, each cone has its own dedicated ganglion cell (1:1). In the periphery, dozens of rods converge onto a single cell with a large receptive field — gaining sensitivity, losing spatial resolution.

The Visual Pathway

Signals leave the retina via the optic nerve and reach the Lateral Geniculate Nucleus (LGN), a six-layered structure in the thalamus. The LGN is organized into magnocellular layers (fast, achromatic, motion-sensitive — fed by parasol ganglion cells) and parvocellular layers (slower, color-sensitive, high spatial frequency — fed by midget ganglion cells). From the LGN, signals project to V1 (primary visual cortex) in the occipital lobe, where orientation-selective neurons (Hubel & Wiesel, 1962) decompose the image into local features. Higher areas — V2, V4 — process increasingly complex properties: contours, color constancy, object recognition.

Eccentricity and the Three Zones

Eccentricity is the angular distance from fixation, measured in degrees of visual angle. The visual field is conventionally divided into functional zones: the fovea (0–2°, highest acuity, cone-dominated), the parafovea (2–5°, where crowding begins to impair recognition), and the periphery (>5°, rod-dominated, coarse spatial pooling). The cortex devotes disproportionate surface area to the fovea — a property called cortical magnification (M-scaling). This is why a letter at 10° eccentricity must be ~5x larger than a foveal letter to be equally legible.

Stage 1: The LGN Gate

"The LGN is not a relay station. It's a gate."
— Sherman & Guillery, The role of the thalamus in the flow of information to the cortex (2002)

The Lateral Geniculate Nucleus sits between retina and cortex. For decades, neuroscientists treated it as a simple relay — signals pass through unchanged. Sherman and Guillery showed it's actually an attentional gatekeeper: top-down signals from cortex modulate what gets through. If you've built information retrieval systems, this is a familiar pattern — a filter stage that decides what enters the processing pipeline at all, before any expensive downstream computation. In the shader, this gating is instantiated as two binary switches and a multiplication. The signal either passes or it doesn't.

Structure Masking

The structure map is packed into a standard RGB texture — a pragmatic reuse of an existing GPU data structure. Each pixel encodes three signals as color channels, which the shader reads as vec3 components. No custom buffer formats or additional render passes needed — the GPU's texture sampling hardware does the interpolation for free.

The LGN gate needs to know what kind of content occupies each region of the page — and how that knowledge is represented constrains what operations are natural downstream. A content analysis pass (content-analysis.js) scans the DOM and encodes three signals into a structure map:

When the LGN stage encounters low density — empty whitespace between content — it suppresses the downstream simulation entirely. The biological LGN doesn't gate blank space to cortex; the simulation mirrors this by zeroing the suppression factor for regions with no visual content to process.

// LGN Structure Masking — lines 319–323 of peripheral2.frag
if (config.lgn_use_structure_mask) {
    if (u_has_structure > 0.5 && signal.density < 0.1) {
        signal.suppressionFactor = 0.0;  // Skip peripheral simulation for empty space
    }
}

Saliency Gating

The saliency map (edge detection + color contrast, temporally smoothed; pop-out reference) tells the LGN what deserves processing bandwidth. This is the same problem biology solves: the retina captures ~107 bits/sec but the optic nerve transmits ~106. The visual system doesn't "degrade then un-degrade" — it selectively allocates limited processing resources to high-value input. Our simulation mirrors this: salient regions receive up to 70% more of the rendering budget (reduced peripheral filtering), while low-saliency regions are processed cheaply.

LGN Suppression Factor
$$\text{suppression} = \text{suppression} \times \text{lerp}(1.0, \, 0.3, \, \text{saliency})$$

At saliency=1.0, the suppression factor drops to 0.3 — the pipeline passes most of the original signal through. At saliency=0.0, full peripheral filtering applies. Both biology and simulation share the same driver: compute demand management.

Dashboard with foveal focus at center
fovea para periph
Dashboard — center focus. The fovea reveals full detail at gaze point. Peripheral regions are filtered through the full LGN→V1→V4 pipeline.
Dashboard saliency map overlay
Saliency map. Bright = high saliency = more processing bandwidth allocated. Text blocks, faces, and colored status badges demand the most resources.

Stage 2: V1 Feature Extraction

V1 orientation selectivity columns — each cortical column responds to a specific edge orientation
V1 orientation columns. Each cortical column responds to edges at a specific angle — 0°, 30°, 60°, 90°, 120°, 150°. This is how the brain decomposes the visual signal into oriented features (Hubel & Wiesel, 1962).

After the LGN gate, the signal that survives enters V1 — and here the nature of the representation changes fundamentally. Primary visual cortex decomposes the signal into features: orientation, spatial frequency, motion. In the periphery, these features are extracted at progressively lower resolution — receptive fields grow larger, features crowd into each other, and positional certainty degrades.

Scrutinizer's V1 stage implements three key peripheral phenomena: positional uncertainty (domain warping), crowding (feature scrambling), and spatial pooling (resolution loss). The LGN's suppression factor modulates all of them — up to 25% attenuation for salient content.

Domain Warping: Positional Uncertainty

In the periphery, you can detect that something is there but not where exactly it is. Scrutinizer simulates this with fractal noise displacement — two octaves of Simplex noise at 150x and 300x frequency, with an anisotropic 2x horizontal bias (horizontal crowding is more severe than vertical in human vision).

Fractal Warp Displacement
$$\vec{d}_{\text{warp}} = \Big(\text{snoise}(\vec{uv} \cdot 150) + \tfrac{1}{2}\,\text{snoise}(\vec{uv} \cdot 300)\Big) \cdot 0.0024 \cdot s \cdot I$$

Where $s$ = distortion strength (eccentricity-scaled) and $I$ = intensity. The 0.0024 constant maps perceptual data: at the foveal boundary, displacement is ~2 pixels. At the far periphery, ~24 pixels — a 12x range matching empirical crowding zones.

Crowding: The Fundamental Limit

"Crowding is not blur." — Denis Pelli (2008). You can detect a letter in your periphery. You cannot identify it when flanked by other letters. The bottleneck isn't resolution — it's obligatory feature integration across the receptive field.

This is arguably the most important phenomenon in peripheral vision — and it's an information-theoretic one (crowding reference stimuli, Bouma spacing). Crowding determines not what you can't see, but what you can't separate. Adjacent features are compulsorily pooled — averaged, texture-ified — once they share a receptive field. The information isn't gone; it's been irreversibly mixed. The individual signals are still in there, but the representation can no longer distinguish them. Bouma's law quantifies the critical spacing:

Bouma's Law — Critical Spacing
$$s \approx 0.5 \cdot \phi$$

Where $s$ = minimum spacing to avoid crowding and $\phi$ = eccentricity from fixation. At 10° eccentricity, targets must be ≥5° apart to be individually identified. This linear scaling with eccentricity mirrors cortical magnification — it's the same constraint expressed differently.

The shader implements two crowding modes:

// V1 Discrete Scramble — lines 403–433 of peripheral2.frag
float scrambleZone = smoothstep(parafovea * 1.0, parafovea * 1.5, dist);
vec2 cellID = floor(uv * cellFreq);
vec2 jitter = hash22(cellID) - 0.5;
vec2 throwDist = vec2(0.008, 0.0016);  // 5:1 H:V ratio
discreteScramble = jitter * throwDist * scrambleZone²;
Article page with foveal focus
fovea para periph
Article — center focus. V1 crowding filters paragraph text in the periphery while headlines receive more bandwidth (larger receptive fields).
E-commerce page with product image focus
fovea para periph
E-commerce — product focus. When you fixate on the product image, the price tag, CTA button, and reviews are filtered through V1 scrambling — a visualization of the feature pooling that constrains peripheral object recognition.

Saliency-Guided Pixelation (Mode 3)

A more utilitarian V1 mode: content is quantized into blocks whose size depends on both eccentricity and the LGN's saliency signal. High-saliency regions get smaller blocks (more detail preserved). The saliency metric is discretized into 4 steps to prevent gradient artifacts:

float steppedMetric = floor(saliency * 4.0) / 4.0;
float blockSize = mix(maxBlock, minBlock, steppedMetric);

Toward Mongrel Textures: DoG Band Decomposition

By the time the signal reaches the periphery, what does the representation actually contain? Rosenholtz et al. (2012) answered this precisely: summary statistics — local texture-like representations that preserve aggregate properties (mean color, dominant orientation, spatial frequency distribution) while discarding individual feature identity. Think of it as a lossy compression where the encoder is the visual system itself. Their mongrel images, synthesized to match these statistics, are perceptually indistinguishable from the original in the periphery. The catch: synthesizing mongrels takes >60 seconds per frame, even with 2026 prosumer hardware.

The shader's DoG band decomposition is a real-time approximation of this idea. Rather than computing full texture statistics, it decomposes the image into spatial frequency bands — each corresponding to a different scale of visual structure — and attenuates them independently based on eccentricity. This mirrors how retinal ganglion cells act as bandpass filters with center-surround receptive fields whose size scales with eccentricity (Curcio et al., 1990). Uniform MIP blurring treats all frequencies equally; the DoG decomposition lets fine detail drop out before coarse structure, matching the biological attenuation profile.

The Approximate Laplacian Pyramid

A MIP chain (from Latin multum in parvo, "much in little") is a series of progressively downsampled copies of a texture that GPUs generate automatically. Level 0 is the full-resolution image; level 1 is half-resolution; level 2 is quarter, etc. GPUs use these for texture filtering at different distances. Hardware MIP generation uses box or bilinear filtering — not the Gaussian convolution of a true Laplacian pyramid (Burt & Adelson, 1983). The shader repurposes this existing hardware feature: subtracting adjacent MIP levels produces an approximate Laplacian pyramid, where each level captures a band of spatial frequencies with some spectral leakage between bands.

The shader decomposes the hardware MIP chain into an approximate Laplacian pyramid — each band captures a spatial frequency range:

Band 0
1–2 px: serifs, thin strokes
Band 1
2–4 px: letter bodies, small icons
Band 2
4–8 px: words, UI elements
Band 3
8–16 px: buttons, layout blocks
Residual
DC: overall color & luminance (always preserved)
// 8 Half-Octave DoG Bands — peripheral2.frag
// 9 MIP levels at LOD 0.0, 0.5, 1.0, ... 4.0 (half-integer = hardware trilinear)
vec4 mip[9];
mip[0] = textureLod(tex, uv, 0.0);
mip[1] = textureLod(tex, uv, 0.5);
// ... mip[2]–mip[7] at LOD 1.0, 1.5, 2.0, 2.5, 3.0, 3.5
mip[8] = textureLod(tex, uv, 4.0);

vec4 band[8];
band[0] = mip[0] - mip[1];  // ~5.66 cpd: serifs, fine detail
band[1] = mip[1] - mip[2];  // ~4.0 cpd:  thin strokes
// ... band[2]–band[7]: √2-spaced down to ~0.5 cpd (layout blocks)
// residual = mip[8]         // ~0.35 cpd: DC, always preserved

M-Scaling: Cortical Magnification per Band

Each band has its own cutoff eccentricity — the distance from the fovea where that frequency band is fully attenuated. With 8 half-octave bands, the cutoffs follow $\sqrt{2}$ spacing, matching the $E_2$ scaling of cortical magnification (validated against published contrast sensitivity data — see report):

Per-Band Cutoff Eccentricities
$$c_k = E_2 \times (2^{k/2} - 1) \qquad k \in \{1, 2, \ldots, 8\}$$

Where $E_2$ is the half-resolution eccentricity — the point where spatial resolution drops to half its foveal value. In the shader: u_dog_e2, user-adjustable. Odd-indexed cutoffs ($c_1, c_3, c_5, c_7$) match the original 4-band anchor values.

Each band's weight is computed via smoothstep rolloff, with a transition width controlled by u_dog_sharpness. At sharpness=0, the transitions are wide and gradual (biological). At sharpness=1, they're narrow and crisp (engineering utility mode).

// Per-band weights via smoothstep rolloff
float w[8];
for (int k = 0; k < 8; k++) {
    w[k] = 1.0 - smoothstep(c[k] - c[k]*trans, c[k] + c[k]*trans, normEcc);
}

// Reconstruction: residual + weighted bands
result = mip[8];
for (int k = 0; k < 8; k++) { result += band[k] * w[k]; }
result = clamp(result, 0.0, 1.0);

What This Means in Practice

Close to the fovea, all 8 bands survive — full detail. In the parafovea, the finest bands (serifs, thin strokes) drop out first. You can still read words but individual letterforms lose their character. Further out, letter-body bands fade — shapes become blobs. In the far periphery, only the coarsest bands and the DC residual remain: you see buttons and layout blocks but not their labels. With half-octave spacing, each transition step is smaller than the old octave-spaced system — the degradation gradient is smoother.

This cascade — fine detail filtered first, coarse structure passing through longest — is how the biological system manages its bandwidth, and it's how the shader manages it too. The signal isn't uniformly blurred; it's selectively attenuated by frequency band, approximating the frequency-selective loss that Rosenholtz's summary statistics model predicts, at real-time rates.

Article structure map
Structure map — article. The content analysis feeds the LGN. Red channel = text rhythm, Green = density, Blue = element type. DoG band weights are modulated by this structure.
Techmeme with foveal focus
fovea para periph
Techmeme — center focus. Dense news layout. DoG preserves headline structure while filtering individual link text. The data table density is exactly the kind of content where DoG outperforms uniform blur.

Fidelity Gap: What DoG Misses

The DoG decomposition captures frequency-selective attenuation but not the full mongrel texture model. True summary statistics include orientation distributions, phase correlations, and cross-scale feature conjunctions that our isotropic bands discard. Adding oriented DoG filters — even just horizontal/vertical — would double the texture lookups but could capture the anisotropy of real V1 receptive fields (Gabor-like, phase-sensitive). The gap between our real-time approximation and a full Rosenholtz-style synthesis is where future work lives.

Stage 3: V4 Color & Aesthetics

Visual area V4 processes color, shape, and object recognition. In the periphery, color perception shifts dramatically — the transition from cone-dominated (foveal) to rod-dominated (peripheral) processing alters hue sensitivity, saturation, and spectral response (per-channel chromatic decay validated — see report). The shader's V4 stage simulates these shifts using Oklab color space, rod-spectrum tinting, and selectable aesthetic rendering modes.

Oklab: Perceptually Uniform Desaturation

A color space is a coordinate system for describing colors. RGB maps to hardware (red/green/blue light intensities) but doesn't correspond to human perception — equal RGB steps don't look like equal brightness steps. Perceptually uniform spaces like Oklab are designed so that equal mathematical distances correspond to equal perceived color differences. This matters when you need to smoothly remove color: in RGB, desaturation causes unwanted hue shifts. In Oklab, you can zero out the chrominance channels (a, b) and get clean, hue-stable grayscale.

A critical design choice. If you desaturate in RGB or HSL, hue shifts as saturation drops — warm colors skew yellow, cool colors skew cyan. Oklab (Ottosson, 2020) separates lightness (L) from chrominance (a, b) in a perceptually uniform space. Killing a and b channels produces clean grayscale without hue contamination.

Oklab Desaturation
$$L_{ab} \cdot \begin{pmatrix} 1 \\ 1 - f \\ 1 - f \end{pmatrix} \quad \text{where } f = \text{smoothstep}(r_{\text{fovea}}, \, r_{\text{fovea}} \cdot k, \, d)$$

$f$ = desaturation factor, ramping from 0 at the foveal edge to 1 at the ramp end. Lightness is preserved. Only chrominance fades. The "Red Kill Switch" (line 625) selectively suppresses the red-green axis first — matching the Purkinje shift where reds darken before blues in peripheral/scotopic vision.

The Five Aesthetic Modes

Mode u_v4_style_id Biology / Design Intent
Usability 0 High-Key Ghosting — eigengrau base, red kill switch. Best for UX evaluation. Shows what users actually perceive in the periphery.
Biological 1 Purkinje Darkening — reds darken, global vignette dims peripheral luminance. The most scientifically accurate mode.
Frosted 2 Frosted glass aesthetic. For presentations where you want peripheral blur without biological artifacts.
Wireframe 3 Edge-detected (Sobel operator) wireframe rendering. Shows structure without surface.
Cyberpunk 4 Saturation boost + halftone dot pattern. For visual storytelling and demos.

Magnocellular Pathway: Luminance Contrast Preservation

The visual system processes information through parallel streams from the retina onward. Parasol ganglion cells (large, fast) feed the magnocellular (M) layers of the LGN (layers 1–2), carrying luminance contrast and motion signals. Midget ganglion cells (small, slower) feed the parvocellular (P) layers (layers 3–6), carrying color and fine spatial detail. These streams remain largely segregated through V1 and into higher cortical areas — the M-stream drives motion perception (dorsal/where pathway), while the P-stream drives object recognition (ventral/what pathway). Livingstone & Hubel (1988) established this segregation; it explains why you can detect motion in far peripheral vision where color and form have long since degraded.

The magnocellular (M) pathway — fast, achromatic, motion-sensitive — preserves luminance contrast even as the parvocellular (P) pathway loses color and detail. The shader samples the clean (undistorted) image, computes a luminance ratio, and blends it back into the simulated output. This is why you can still detect edges and motion in your far periphery even when you can't see color or detail.

// Magnocellular Pathway — lines 555–570 of peripheral2.frag
float cleanLuma = dot(cleanSample, vec3(0.299, 0.587, 0.114));
float distortedLuma = dot(col, vec3(0.299, 0.587, 0.114));
float lumaRatio = cleanLuma / max(distortedLuma, 0.01);
float preservation = mix(0.6, 0.1,
    smoothstep(0.0, parafovea - fovea, eccentricity));
col *= mix(1.0, lumaRatio, preservation * contrastRamp);
Techmeme structure map
Structure map — Techmeme. The structure analysis that feeds the LGN gate. Dense news sites reveal the rhythm/density/type decomposition that drives the entire pipeline.
Techmeme saliency map
Saliency map — Techmeme. Edge detection + color contrast in Oklab space. Bright regions receive more processing bandwidth through the LGN gate.

Project Trajectory — 374 commits in 98 days, three shader reverts in 30 minutes, and the gap between biological accuracy and perceptual plausibility. Read the development arc →

The Full Mapping: Biology → GLSL

Each row below maps a biological phenomenon to its shader implementation, with the controlling parameter. For full citations and methodology, see the scientific literature review.

Biological Phenomenon Shader Implementation Key Uniform / Constant
Cone density (1:1 foveal wiring) Full-resolution sampling at eccentricity = 0 u_foveaRadius
Rod density (100:1 convergence) MIP pyramid / DoG band attenuation u_dog_e2, u_dog_sharpness
Ganglion cell receptive fields Difference-of-Gaussians band decomposition band0–band3, c0–c3
Cortical magnification (M-scaling) Per-band cutoff eccentricities, geometric 2x c_k = 0.3 · 2^k · E2
LGN attentional gating Saliency-modulated suppression factor u_lgn_use_saliency_gate
V1 crowding (Pelli 2008) Fractal noise warp + discrete scramble u_v1_distortion_type
Positional uncertainty Simplex noise domain warping, 2x H bias fractalWarp.x *= 2.0
Purkinje shift Red-green axis kill in Oklab lab.y *= (1.0 - fade)
Rod spectral sensitivity (505nm) Cyan tinting in far periphery rodColorLab vec3(L, 0.0, -0.05)
Magnocellular luminance preservation Clean/distorted luma ratio blending contrastPreservation mix(0.6, 0.1)
Saccadic suppression Velocity-gated jitter modulation u_velocity
Eigengrau (dark adaptation) Background color at 0.96 lightness in Oklab vec3(0.96 * lab.x, 0.0, -0.05)

Mobile: The Smaller Fovea

On mobile, the fovea subtends the same visual angle, but the screen is held closer and the viewport is denser. The simulation reveals something counterintuitive: much of the content on a phone screen sits in the parafoveal zone — not the far periphery, but the transitional region where feature integration begins to degrade. This is where crowding first bites.

Article on iPhone 14 viewport
fovea para periph
iPhone 14 — article. The same article page at mobile scale. Much of the content sits in the parafoveal zone, where feature integration starts to blur individual elements.
Dashboard on iPhone 14 viewport
fovea para periph
iPhone 14 — dashboard. Data-dense UIs on mobile. The fovea reveals one card; most content falls into parafoveal territory.
Article on iPad Air viewport
fovea para periph
iPad Air — article. Tablet's larger viewport provides a middle ground. More foveal real estate, but still significant parafoveal load.
Dashboard on iPad Air viewport
fovea para periph
iPad Air — dashboard. The data table density makes this an ideal test case for the DoG band decomposition — rows fade by spatial frequency, not uniformly.

Select References

Further Reading

About the Author

Andy Edmonds studied cognitive science at Carnegie Mellon, Georgia Tech, and Clemson, then left academia in 1995 for what became two decades of machine learning at scale — shipping recommendation systems, search relevance, and information retrieval products at eBay, Adobe, Microsoft, and Meta. The through-line across all of it is the same question that drew him to cognitive science in the first place: how does information flow through a system, and what happens to the signal at each stage?

Scrutinizer is that question applied to the visual system. The retina decomposes. The LGN gates. V1 pools. V4 recolors. Each stage transforms the signal in specific, modelable ways — and those transformations map cleanly onto GPU operations. The shader isn't a metaphor for the biology. It's an implementation of the same information-processing architecture, running on different hardware.