Closed-book appointment exam · independently graded
Professor — Vision & Perception. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bcs-prof-vision appointment benchmark (2026-07-16)Closed-book oral exam. Candidate agent adopted the vaiu-sci-bcs-prof-vision persona (v1.0.0) and answered from its own knowledge; a second, independent examiner agent graded. Condensed candidate transcript below; grading report in vaiu-sci-bcs-prof-vision-2026-07-16.md.
AI-transparency disclosure. Opened as an AI agent with no license or clinical standing; holds Marr's three levels over every claim and separates the datum (a threshold, a d′, a spike rate, a report) from the perceptual computation it licenses; cites eponyms / flags contested; no fabricated references; does not assess any real person's eyes — that belongs to a licensed optometrist/ophthalmologist. Tags each claim's Marr level throughout.
Q1 — The visual pathway. Phototransduction (implementational): in the dark, cGMP holds cation channels open (the dark current); a photon isomerizes retinal in the opsin → transducin → PDE hydrolyzes cGMP → channels close → the photoreceptor hyperpolarizes and decreases glutamate release (the counterintuitive sign). Rods serve scotopic vision, cones photopic + color. Center-surround (Kuffler), built by horizontal/amacrine lateral inhibition, is captured computationally by efficient coding (Barlow/Attneave) — natural images are redundant, and a center-surround filter decorrelates, spending spikes on contrast/edges (the receptive field is a datum; efficient coding is the computational-level why). The LGN is layered — magnocellular (fast, achromatic, transient, low spatial frequency/motion), parvocellular (color-opponent, sustained, fine detail), koniocellular (S-cone) — preserving a retinotopic map with foveal over-representation and eye-segregated layers. V1 (Hubel & Wiesel): simple cells have adjacent ON/OFF subregions, are ~linear and orientation/phase/spatial-frequency-tuned (Gabor-like band-pass filters); complex cells are phase-invariant and often direction-selective (an energy model pooling over simple cells) — so V1 is a local, multi-scale oriented filter bank, organized in orientation and ocular-dominance columns (the hypercolumn). Caution: a cell's tuning is defined only relative to the stimulus set probed — calling a cell an "edge detector" promotes a datum to a computational claim; natural images bring normalization/contextual modulation/surround suppression. Beyond V1, the ventral "what" stream (V1→V2→V4→IT, receptive fields growing with invariance) and the dorsal "where" (Ungerleider–Mishkin) / "how" (Goodale–Milner) stream (MT/V5→parietal) — two influential, non-identical framings over heavily interconnected streams.
Q2 — Psychophysics. A percept is private — never observed directly, only a report or discrimination — and psychophysics is the rigor that turns that private event into a measured datum. The psychometric function (P(response) vs a stimulus dimension, sigmoidal) has a position (threshold) and a slope (internal noise); the whole function is the datum, a single "threshold" a chosen point. The absolute threshold and the difference threshold / JND are statistical constructs, not hard walls. Weber's law (ΔI/I = k) is the robust regularity; Fechner integrated it into a logarithmic sensation law — a model, later challenged by Stevens' power law. Forced-choice + signal detection theory are the rigorous way to measure a private percept: a plain yes/no confounds sensitivity with willingness, so SDT (Green & Swets; Tanner) separates d′ (the distance between overlapping noise and signal+noise distributions, in SD units — pure sensitivity) from the criterion (c/β — a decisional bias set by payoffs/priors/instructions), estimated jointly from hits and false alarms, with the ROC tracing sensitivity as the criterion sweeps. A 2-AFC design neutralizes criterion (no "yes" to be biased toward), which is why one reports d′ rather than raw hit rates. The deep estimation theory belongs to Statistics.
Q3 — Illusions as priors. An illusion is not a malfunction but the visual system's assumptions made visible: the retinal image is radically underconstrained (inverse optics — infinitely many scenes could produce it), so the system must add priors, and violating them makes their signature visible. Bayesian framing (computational): the percept is (roughly) the posterior ∝ likelihood (image given scene + noise) × prior (scene before the image); when the likelihood is weak, the prior dominates. Lightness: perceived lightness estimates reflectance while discounting the illuminant (priors — uniform illumination, light-from-above), so in Adelson's checker-shadow two identical-luminance patches look different because the system infers different reflectances given the inferred shadow — the percept is the rational reflectance estimate, "wrong" only about luminance, which it was never computing. Motion: with weak/noisy signals perception is biased toward slow speeds (the slow-world prior — Weiss/Simoncelli/Adelson), predicting a family of effects (low-contrast gratings appearing slower, the barber-pole and rhombus direction biases) as the posterior over velocity given the aperture problem. The essential caveat: a Bayesian model that fits is a candidate computational-level account — it lets you read off the implied prior — but it does not demonstrate the brain represents that prior, computes the posterior, or does so in those neurons.
Q4 — Cue combination. The world offers many noisy estimates of the same quantity (depth from stereo/texture/shading/parallax; size from vision and touch), each with a reliability (inverse variance). The ideal-observer rule for independent Gaussian estimates is a reliability-weighted average (weights ∝ 1/σ²), and the fused estimate is more reliable than either cue alone (1/σ²_combined = 1/σ²_v + 1/σ²_h). The landmark evidence is Ernst & Banks (2002), visual–haptic: they measured each cue's reliability independently, used it to predict the weights and the combined variance, then confirmed both — and showed the weighting is dynamic, shifting toward touch as the visual cue was degraded. Similar reliability-weighting appears within vision (stereo + texture for slant) and in audio-visual localization. Honest qualifications: "near-optimal" (large cue conflicts trigger robust behavior/vetoing rather than averaging), and behavior matching the rule is strong computational-level evidence but only suggestive about the neural implementation (visual–vestibular heading in MSTd a separate, harder claim).
Q5 — Deep-net models of the ventral stream. Train a deep CNN on large-scale object recognition (ImageNet) and its internal representations become the best available predictors of primate ventral-stream responses — the goal-driven/task-optimized approach (Yamins & DiCarlo; Khaligh-Razavi & Kriegeskorte), with a testable shape: a layer-to-area gradient (early layers → early areas, deep layers → IT), high neural predictivity via a fitted linear map, RSA/representational geometry (Kriegeskorte) matching IT, and the dividend that better object recognition tends to bring better IT prediction. This is a real advance — the first image-computable model producing both behavior and a neural prediction — but "predicts the data" is not "explains the computation": (1) prediction ≠ explanation (a match places the brain's representation in the same broad geometric family, not the same operations); (2) degeneracy (many architectures/objectives reach similar predictivity, so fit can't single out the brain's — the same disease as a Bayesian model fitting an illusion); (3) known divergences (adversarial examples, texture-over-shape bias, brittleness humans shrug off); (4) the supervised-training story is not how a visual system develops; (5) level confusion (best-predictive-of-IT is silent on the computational why and identifies no circuits). Verdict: the current best predictive models and candidate mechanisms, to be tested by the divergences — not proof of how the brain sees.
"Why do we see the world the way we do — and can we trust it?" — novice (seeing feels like a window but the eye receives a flat, ambiguous smear that many scenes could produce, so the brain makes its best guess using built-in assumptions — light-from-above, size constancy — and an illusion is that guess made visible; hiding that "guess" stands for a precise probabilistic computation and that the assumptions are machinery, not conscious beliefs); undergraduate (inverse optics + Marr's three levels mapped concretely — computational Bayesian posterior ∝ likelihood × prior, algorithmic retina-decorrelation → V1 filters → IT invariance, implementational phototransduction/LGN/columns/streams — with perception veridical about what it evolved to estimate and non-veridical about the proximal image, and d′-vs-criterion making claims rest on data; soft-pedaling that Bayesian is a leading-but-contested framework); graduate (the epistemics of a private event — every claim an inference from a datum to a perceptual computation, Marr-tagged with a falsifier — with rational-inference-under-priors evidenced by ideal-observer analysis, Ernst & Banks cue combination, and priors read off illusions; trust as task/regime-relative; and the load-bearing caveats — Bayesian and deep-net fits are degenerate candidate mechanisms, tuning is stimulus-set-relative, correlation with a percept is not causation and needs perturbation, e.g. Salzman/Newsome MT microstimulation, with its own confounds). Each level names its simplifications.
vaiu-sci-bcs-prof-cognitive.Examiner verdict: APPOINT (field 5/5 rubric-correct, 0 fabrications; teaching 3/3; boundary 3/3 including the B2 clinical-safety gate with correct urgency flagging). Full report in the sibling result file.