Generic views, Bayesian flexibility, and a falsifiable invariance test

Definitions and mathematical statement

Nature of Chapter G. This is a discussion after Ken Nakayama's talk, not a formal conjecture. Dana asks whether generic views agree with Bayes. The reply calls Bayesian priors stochastic/nonergodic, invokes Kanizsa diagrams and criticizes Bayesian unconstrainedness. Other comments ask whether generic views themselves are priors, mention Bill Freeman's Bayesian treatment, note ambiguous figures (speaker marked "Jeff?"), and ask about motor experience. The discussion contrasts impossible-triangle perception with explicit knowledge, asks about neurobiological explanation and models versus theory, records "Reductionism is metaphysical," and emphasizes developmental kinetic-depth/stereo trajectories and perception--motor interaction. Its optimistic/pessimistic-camp remark is commentary, not a mathematical assertion.

The aid's identified mathematical issue is whether generic-view accounts are distinguishable from unconstrained Bayesian rationalizations. The following finite-model statements are introduced by that aid; they are not equations or posed theorems in the transcript.

Reported prior-absorption criterion. Let $\mathcal H,\mathcal I$ be finite hypothesis/image sets, with positive prior $\pi(h)$, positive specified likelihood $L(i\mid h)$ and positive genericity multiplier $G(h,i)$. The posterior proportional to $\pi(h)L(i\mid h)G(h,i)$ can use the same likelihood and absorb $G$ entirely into one image-independent object prior exactly when \[ G(h,i)=a(h)b(i) \] for positive functions $a,b$. Then the new prior is proportional to $\pi(h)a(h)$. Full positivity makes this equivalent to all cross-product identities $G(h,i)G(h',i')=G(h,i')G(h',i)$. With zeros, the criterion must be interpreted on the actual likelihood support; the supplied summary does not state those cases.

Reported arbitrary-table rationalization. Given any conditional table $Q(h\mid i)\geq0$ with $\sum_hQ(h\mid i)=1$, choose any strictly positive image evidence $e(i)$ summing to one. Define \[ P(h,i)=e(i)Q(h\mid i),\qquad \pi(h)=\sum_i e(i)Q(h\mid i),\qquad L(i\mid h)=\frac{e(i)Q(h\mid i)}{\pi(h)}\quad\text{if }\pi(h)>0. \] If $\pi(h)=0$, every $Q(h\mid i)=0$ and $L(\cdot\mid h)$ may be any probability vector. These choices give a normalized prior/likelihood model reproducing $Q$ on every image, since $e(i)>0$. The aid states that every joint law with full image support is recovered by choosing its own image marginal as $e$. For a general joint law, conditional probabilities are fixed only where the image marginal is positive; zero-evidence images remain version-arbitrary. Zero-prior hypotheses and zero-evidence images are therefore different phenomena.

Reported common-prior intervention test. Suppose environments $v$ implement specified viewpoint interventions and supply fixed positive generic-view kernels $K_v(i,h)$, incorporating any declared nuisance-view integration and genericity factors. Given positive observed posteriors $Q_v(h\mid i)$, the existence of a common positive prior with \[ Q_v(h\mid i)=\frac{\pi(h)K_v(i,h)}{\sum_{h'}\pi(h')K_v(i,h')} \] is equivalent to constancy across images and environments of every corrected odds ratio \[ \frac{Q_v(h\mid i)}{Q_v(h'\mid i)}\frac{K_v(i,h')}{K_v(i,h)} =\frac{\pi(h)}{\pi(h')}. \] The kernels and interventions must be specified independently; arbitrary refitting removes this restriction. The summary reports a worked two-view surface family with a prior-free ranking-reversal prediction, but gives no numerical kernels or reversal values, so that example is not reconstructed here.

The aid attributes to Nakayama--Shimojo a generic-sampling account as a degenerate Bayesian form neglecting object priors, and to Freeman viewpoint-prior integration yielding a marginal image likelihood. Full primary references and the transcript's nonergodicity definitions are absent. These conditional finite-model audits do not settle the discussion's philosophical, developmental or neural-mechanism questions. The registry says open/OPEN-TRIAGE, while the attempt itself says context-only/not-a-problem; both scopes are retained.

Short English statement

The source is a 2003 discussion about generic views, Bayes and neural explanations. The aid supplies finite-model tests: when a genericity factor can become an object prior, why an unrestricted conditional table always has a Bayesian rationalization, and how fixed viewpoint interventions can test a common-prior restriction. These formalizations do not turn the transcript into a solved formal conjecture.

Sources

• Local UnsolvedMath registry, permanent ID $20000864$, catalogue entry; canonical aim-biology-notes.json, notes[28], tagged section.
• AIM, "Inference and Prediction in Neocortical Circuits," Chapter G, "Helmholtz Inference in Early Vision Areas," discussion after Ken Nakayama's talk: https://aimath.org/WWN/brain/brain.pdf. Supplied verification dates the hard-copy web-page version October 24, 2003 and locates it at PDF indices 7--8, printed pages 8--9; it contains no talk-summary equations or formal conjecture.
• Transcript names Dana, Jeff (one attribution uncertain), and Bill Freeman; the aid also names Nakayama--Shimojo and Freeman's primary lineage. No full paper titles or bibliographic entries are supplied, and no independent source visit is claimed.
• Research attempt 1, context_only, run status not_a_problem, internal rigor proved, low novelty confidence. Explicitly a machine-generated research aid, not peer reviewed; independently verify before citation. Original literature field empty.