COLLEGIUMCOLLEGIUM
Hall of Shoulders

AI Reasoning

David Marr

David Marr (1945-1980) was a British neuroscientist and vision theorist whose posthumous *Vision* (1982) reframed how complex information-processing systems are explained. His central, enduring contribution is a methodological insight, not a single algorithm: that any system that processes information must be understood at three logically distinct levels, and that confusing those levels is the most common way to produce theories that are wrong, untestable, or explanatorily empty. This dossier applies Marr's framework as a *review lens* for space-domain dissertation candidates: it is a discipline for asking whether a proposed space-system capability is well-posed before asking whether it is well-built.

Built

Sources

49

Primary + secondary

Citations

0

ARGOS-tracked

FTS5 Chunks

49

Retrieval index

Councils

0

Memberships

Review Lens

Adversarial questions for candidates

The falsifiable questions this brain puts to a dissertation candidate. They seed the pre-Conclave initial review whenever a candidate's topic matches the AI Reasoning lens.

  1. 1

    Computational theory test. State, in one sentence, the function your space

  2. 2

    Level-separation test. Identify which of your contributions live at the

  3. 3

    Representation-justification test. Why is your chosen representation

  4. 4

    Appropriateness / objective-function audit. Your optimizer minimizes some

  5. 5

    Level-bridging / feasibility-feedback test. For a flight system, show how

Core Concepts & Space Translation

F1 - The three levels of analysis

*(Marr, Vision, 1982; canonical.)* Every information-processing system admits three complementary levels of description: (1) the **computational theory** - WHAT problem is being solved and WHY this is the right problem (the goal, and the logic of the mapping from input to output, including why that mapping is appropriate given the structure of the world); (2) the **representation and algorithm** - HOW the problem is solved (what representations encode the input and output, and what procedure transforms one into the other); (3) the **hardware implementation** - how the algorithm is physically realized in a substrate. The levels are loosely coupled: one computation admits many algorithms, and one algorithm admits many implementations.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

F2 - Primacy of the computational level

*(Marr 1982; Shagrir, "Marr on Computational-Level Theories," 2010.)* Marr's deepest claim is that the computational theory is the most important and most neglected level, because "the nature of the computations that underlie perception depends more upon the computational problems that have to be solved than upon the particular hardware." Shagrir sharpens this: the computational level has both a **What** (the function computed) and a **Why** (an argument that this function is optimal or appropriate given environmental constraints). A theory that specifies only mechanism, with no account of why the system *should* compute that function, is incomplete.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

F3 - The "Why" / appropriateness argument

*(Shagrir 2010.)* A genuine computational-level theory does not merely list an input-output mapping; it explains why that mapping is the right one for an agent embedded in a particular environment. In space terms: a conjunction-screening rule is not justified by being implementable, but by an argument that it correctly trades miss-distance risk against maneuver cost given the actual statistics of the orbital population.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

F4 - Representations are theory-laden and task-specific

*(Marr 1982; the primal sketch / 2.5D sketch / 3D model progression.)* Marr held that choosing a representation is a substantive theoretical commitment, because a representation makes some information explicit at the cost of others. His vision pipeline moves from a primal sketch (intensity changes, edges) to a viewer-centered 2.5D sketch to an object-centered 3D model - each making explicit what the next stage needs.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

F5 - Level confusion as a diagnostic failure mode

*(Peebles & Cooper, "Thirty Years After Marr's Vision," 2015; Bechtel-tradition mechanistic readings, e.g., Milkowski-style integration, 2020.)* Most flawed information-processing theories fail by conflating levels: explaining behavior directly from hardware (eliminating the algorithmic level), or treating an optimization story as if it specified a mechanism. The corrective is to keep the levels separate and then explicitly connect them.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

F6 - Levels are bridges, not silos (the modern revision)

*(Genkin-style "Connecting levels of analysis in the computational era," 2023; mechanistic- computational integration, 2020.)* Contemporary work argues the levels are most useful when productively *connected* - implementation constraints feed back into which algorithms are feasible, and algorithmic discoveries reshape the computational problem. For space systems, where flight hardware is severely resource-constrained, this two-way coupling is not optional.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.