research-package RP-COMP-005
Visual Scene Construction, Predictive Processing, and Active Perception
Evidence supports a hybrid account in which rapid feedforward processing produces provisional scene and object hypotheses, recurrent and contextual interactions refine them, and active sampling obtains additional evidence. Prediction is important, but predictive coding is not established as the unique neural account.
RP-COMP-005 — Visual Scene Construction, Predictive Processing, and Active Perception
Research State
This canonical repository record integrates the completed execution package into the Composition Science artifact system. The full research journal, evidence registry, hypothesis registry, bibliography, debt register, and completion checklist remain in the source package identified in frontmatter.
Original Question
How does raw sensory input become an actionable understanding of a visual scene, and can this process provide a predictive foundation for Composition Science?
Result
The evidence rejects a strictly linear perception pipeline and supports a hybrid model:
Goals and priors
↓
Rapid parallel scene/object hypothesis
↓
Provisional grouping and hierarchy
↓
Task-relevant uncertainty
↓
Attention and action selection
↓
Eye/head/body/interface action
↓
New high-resolution evidence
↓
Recurrent integration and revision
↺
Key Discoveries
- KD-001: initial scene understanding is parallel and provisional.
- KD-002: rapid recognition does not imply complete understanding.
- KD-003: recurrent processing supports stable conscious perception.
- KD-004: contextual predictions can help or hurt depending on signal and task.
- KD-005: expectation suppression is not general proof of predictive coding.
- KD-006: high-level unconscious integration remains unresolved.
- KD-007: eye movements are part of the computation.
- KD-008: composition guides evidence acquisition.
Hypothesis Outcomes
| ID | Outcome | Confidence |
|---|---|---|
| HY-COMP-033 | Supported after revision: feedforward initiation plus recurrent and active refinement | High |
| HY-COMP-034 | Provisionally supported with limits on unconscious semantic integration | Medium |
| HY-COMP-035 | Supported only as bounded task-relevant uncertainty reduction | Medium |
| HY-COMP-036 | Original rejected; revised to management of useful and unnecessary prediction error | Medium-High |
| HY-COMP-037 | Rapid gist hypothesis supported | High |
| HY-COMP-038 | Recurrent integration provisionally supported | Medium-High |
| HY-COMP-039 | Active sampling supported | High |
Theory Impact
Create TH-COMP-005, replace the linear core diagram with a recurrent perception-action loop, and treat prediction-error minimization as a bounded mechanism rather than a master law.
Candidate laws introduced by the package are LAW-COMP-033 through LAW-COMP-039: feedforward-recurrent complementarity, conditional context, task-relevant uncertainty, controlled violation, fixation-dependent evidence, evidence continuity, and verification cost.
Operational Outputs
Evidence Traceability
The package registers EV-COMP-051 through EV-COMP-060. Load-bearing sources include critical reviews and primary studies on scene/object interaction, masking and recurrence, expectation suppression, context reliability, eye movements, and active vision. Claims must preserve these boundaries:
- prediction as a functional property is not equivalent to predictive coding as a neural mechanism;
- fixation is not comprehension;
- rapid gist is not complete understanding;
- context is conditional rather than uniformly helpful;
- static screenshots do not model the whole natural viewing process.
Next Research
The highest-value successor is RP-COMP-006 — Eye-Movement Guidance, Visual Search, and Fixation Prediction. It should begin with a consolidated theory registry, use primary studies and critical reviews, seek failed replications and alternative models, and produce at least one falsifiable experiment.