Constructive Visual Imagery And Perception

Arthur M. FarleyIJCAI 1975

Abstract

This paper introduces the Visual Imagery and Perception System (VIPS), a computational model designed to simulate human visual perception as an active, constructive process. Developed in 1974 by Arthur M. Farley, VIPS challenges traditional "snapshot" theories by proposing that the brain builds mental images through a sequence of glimpses, integrating these into a symbolic internal representation. Through production systems and a sophisticated memory hierarchy, VIPS successfully models how humans perceive complex forms by testing hypotheses and directing eye movements.

Introduction

In the early 1970s, theories of visual perception were divided between Atomism—the assembly of images from unconscious patches—and Gestalt theory—the immediate recognition of holistic forms. Farley (1974) argued that both were insufficient for computational implementation. Atomism fails to explain temporal integration, while Gestalt theory lacks the specificity required for computer modeling. The core problem addressed is how a stable, complete perception arises from the fovea's limited focus. Farley proposes a "constructive" model where perception is an active, goal-directed cycle of fixation and hypothesis testing.

The VIPS Architecture

The Visual Imagery and Perception System (VIPS), implemented in LISP 1.6, operates on a cycle of fixation, feature extraction, and schematic map updates.

  • Image Chunks: Mental representations are stored as "chunks"—symbolic structures like vertices or lines—organized in doubly linked circular lists.
  • Production Systems: The system utilizes "condition-action" rules. For instance, detecting a right angle might trigger a search for a parallel line to confirm a square.
  • Memory Hierarchy: VIPS employs a memory stack consisting of Short-Term Memory (STM) for the "mind's eye" and Long-Term Memory (LTM) for geometric knowledge.

Experimental Validation

Farley validated VIPS using two primary experimental paradigms:

  1. The Hole-Viewing Task: Subjects identified line drawings by moving a small aperture (3-degree angle) over the image while providing "think-aloud" protocols.
  2. Eye-Tracking: Corneal reflection was used to track actual eye movements during unconstrained viewing of geometric stimuli.
Sequence of glimpses (V1-V9) and verbal protocols from the hole-viewing task, illustrating the constructive identification of a square and triangle.

Key Insights and Results

The research yielded several significant findings regarding the nature of human vision:

  • Perception as Hypothesis Testing: Vision is not passive; it is a process of confirming hypotheses. A detected corner leads to a search for edges to confirm a specific shape.
  • Symbolic Mental Imagery: The "mind's eye" is modeled not as a bitmap but as a network of symbols and spatial relations.
  • Behavioral Matching: VIPS accounted for over 80% of the hole-movements observed in human subjects.
  • Scanpath Integration: The system demonstrated that the initial fixation acts as a "seed," determining the subsequent search strategy and inferences.
A recorded scanpath on a 3D wireframe object, showing the sequence of fixations (1-22) used to construct a mental representation of the complex form.

Limitations and Conclusion

While VIPS represents a milestone in cognitive AI, it had limitations, notably the exclusion of peripheral vision and a focus on simple line drawings without color or texture. However, the model's success in matching human behavioral data suggests that constructive, multi-strategy systems are essential for robust computer vision. Farley’s work remains a foundational argument for seeing vision as an active, intelligent process rather than a mere sensory recording.

Summary updated

BibTeX
@inproceedings{farley1975constructive,
  author = {Arthur M. Farley},
  title = {{Constructive Visual Imagery And Perception}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure