Advanced Topics¶
Visuospatial Reasoning¶
Visuospatial reasoning is reasoning with visuospatial knowledge, which has two components:
Visual — the “what” (objects, identities)
Spatial — the “where” (locations, spatial relations)
A defining characteristic of visuospatial knowledge is that causality is at most implicit. A picture of a fallen cup with a pool of water doesn’t explicitly state causation, but we infer the cup fell and spilled. Visuospatial knowledge enables inferences about causality without representing it directly.
Propositional vs. Analogical Representations¶
Two approaches to handling visuospatial knowledge:
Propositional representations (amodal):
Extract symbolic descriptions from figures (e.g., “triangle, apex right, rotated 90°”)
Manipulate these descriptions using logic or production rules
Divorced from perceptual modality
This is how most AI systems currently work
Analogical representations (modal):
Maintain structural correspondence between the representation and the external figure
Apply affine/set transformations directly (rotation, reflection)
Close to the perceptual modality
Human mental imagery appears to use analogical representations
The same distinction applies to other modalities (e.g., auditory). Open questions in cognitive science: When a melody reminds you of another, are you matching propositional extractions or directly matching analogical representations?
Knowledge Content and Encoding¶
Content |
Encoding |
Example |
|---|---|---|
Visuospatial |
Analogical |
Mental image of cup |
Visuospatial |
Propositional |
“circle at (3,4)” |
Verbal |
Propositional |
Script with slots |
Verbal |
Analogical |
Short mental movie |
Most of KBAI deals with verbal/propositional knowledge. Fully understanding and leveraging visuospatial + analogical representations remains an open challenge.
Galatea: Visuospatial Analogical Transfer¶
Galatea (Jim Davies, Georgia Tech) demonstrated analogical problem-solving using purely visuospatial knowledge without extracting causal patterns.
The Duncker radiation problem: A physician must destroy a tumor with a laser, but the beam at full strength kills healthy tissue. The analogical source is a story of an army decomposing into small groups to converge on a fortress from multiple directions (avoiding mines that trigger on large forces).
Most computational models extract a propositional causal pattern: “if resource can achieve goal but obstacle blocks it, decompose resource and converge from multiple directions.”
Galatea instead transferred the visuospatial structure step-by-step — mapping spatial elements (top road → top body part, fortress → tumor) without abstracting a causal pattern. The causality remains implicit but the problem-solving procedure transfers purely through spatial correspondence.
Archytas: Causal Models from Drawings¶
Archytas (Patrick Yaner, Georgia Tech) extracted causal/functional models from vector graphics engineering drawings.
Process:
Library of known drawings with pre-built hierarchies: line segments → basic shapes → composite shapes (piston, cylinder) → behavioral models → functional specifications
New drawing input → generate line segments and arcs
Map to closest known drawing via analogical matching
Transfer labels up the abstraction hierarchy (shapes → components → behavior → function)
Result: Given a new piston-crankshaft drawing, Archytas assembles a causal model (“linear motion converts to rotational motion”) by analogy to known drawings — extracting causal information from visuospatial representations.
Raven’s Progressive Matrices¶
Multiple computational approaches to Raven’s test use visuospatial representations:
Maithilee Kunda’s VITA system uses purely visual-spatial representations (no propositional extraction)
Keith McGreggor’s system uses fractal representations
Both achieve good accuracy, demonstrating that analogical/visuospatial representations can support complex reasoning without propositional intermediaries.
Systems Thinking¶
The external world consists of systems — heterogeneous interacting components whose interactions produce processes at multiple levels of abstraction, some invisible.
Examples:
Ecosystems: physical, biological, chemical processes interact across scales
Businesses: manufacturing, marketing, delivery units at individual/team/organization levels
Systems thinking = reasoning about invisible properties and complex behaviors of systems; deriving invisible processes from visible structure.
Structure-Behavior-Function Models¶
SBF models capture three levels:
Structure — visible components and their connections (e.g., bulb, switch, battery connected in a circuit)
Behavior — invisible causal processes as state transitions (e.g., switch closes → electricity flows from battery to bulb → bulb converts electricity to light)
Function — the purpose/input-output characterization (e.g., “create light”: stimulus on switch, 0 lumens → 30 lumens)
Key properties:
SBF models are nested — the flashlight has an SBF model, and the bulb within it has its own SBF model
They capture multiple levels of abstraction
They enable diagnosis and design of complex systems by making invisible causal processes explicit
Design Thinking¶
Configuration (routine design): all components are known; the task is finding an arrangement and assigning variable values. Methods include plan refinement, case-based reasoning, model-based reasoning, rule-based reasoning.
Creative design goes beyond configuration — not all parts are known in advance. It involves:
Problem-solution co-evolution (the problem evolves as the solution evolves)
Ill-defined, underconstrained, open-ended problems
IDOL: Learning Design Patterns¶
IDOL (Sam Bhatta, Georgia Tech) performed creative design by learning and transferring design patterns across domains.
Example:
A 1.5V battery circuit produces 10 lumens
Goal: produce 20 lumens, but no 3V battery exists
Solution: connect two 1.5V batteries in series (cascading pattern)
IDOL extracted the general pattern: if a design achieves value X via behavior B1, achieve value 2X by replicating B1 through series composition.
When later given the problem of designing a higher-capacity water pump, IDOL transferred the cascading pattern from electrical circuits — connecting multiple pumps in series. This is analogical transfer of design patterns across domains.
Creativity¶
Defining creativity (one formulation): producing a novel, unexpected, desirable output.
Novelty = newness (making soufflé for 20 people when you’ve only made it for 4)
Unexpectedness = surprise (inventing a dramatically different soufflé recipe)
Processes of Creativity¶
Previously covered in this course:
Analogical reasoning — transferring patterns across domains (water pump from electrical circuits; atomic model from solar system)
Explanation-based learning — creative reuse (using a flower pot as a cup)
Additional processes:
Emergence — drawing three lines produces a triangle; the triangle-ness doesn’t belong to any single line and wasn’t the intent
Re-representation — when the original representation blocks problem-solving, restructure it (e.g., switching from verbal “revolves/rotates” to spatial diagrams to see the solar-system/atom analogy)
Serendipity — a suspended goal connects with an unrelated discovery (Mestral’s Velcro: stuck zipper problem + burrs on dog’s legs under microscope)
Conceptual combination — blending two concepts to produce something new
Can AI Be Creative?¶
Arguments against AI creativity (and their rebuttals):
“Algorithm outputs can’t be novel” → Combinations of algorithms for open-ended problems can produce novel outputs (e.g., design, scientific discovery)
“Given same input, output is always the same” → Output depends on input + method + context/situation; same input in different contexts yields different outputs
“If we can trace the process, it’s not creative” → Creativity can be defined by the output alone (a black box producing interesting music is creative regardless of whether we understand the process)
AI Ethics¶
Key ethical questions in AI:
Economy and society — AI agents replacing human jobs (e.g., assembly robots). Counter: new jobs in designing/maintaining AI. Hard tradeoffs remain.
Military applications — Drones and robot soldiers. Should we build morality into autonomous weapons? If so, what does this reveal about human morality?
Civil rights for machines — At what point do AI agents become sufficiently human-like that we must consider their rights? What are the criteria for considering machines equal to humans?
These are open questions without easy answers, increasingly relevant as AI capabilities advance.