The Structural Affinity Method for Solving the Raven's Progressive Matrices Test for Intelligence

Snejana Shegheva, Ashok K. GoelAAAI 2018 · DOI

Abstract

This paper presents the Structural Affinity method, a novel computational approach for solving Raven’s Progressive Matrices (RPM) using Markov Random Fields. By modeling image relationships as statistical affinities rather than rigid symbolic rules, the method bridges the gap between raw pixel data and abstract logical reasoning. Evaluated on the Standard Raven Progressive Matrices, the model achieved a 91.67% success rate, demonstrating performance comparable to or exceeding state-of-the-art symbolic and visual models.

Introduction

The Raven’s Progressive Matrices (RPM) test is a cornerstone of non-verbal intelligence assessment. It requires subjects to identify a missing piece in a geometric pattern, typically arranged in a 2x2 or 3x3 grid. Solving these problems necessitates the induction of abstract rules—such as rotation, addition, or constancy—directly from visual stimuli.

Historically, computational approaches to RPM have been divided into two paradigms: analytic models that rely on hand-coded symbolic descriptions, and gestalt models that operate on raw pixels but often fail to capture complex logical structures. The Structural Affinity method aims to bridge this divide by treating logical deduction as statistical inference over graphical models.

The Structural Affinity Method

The core of the proposed approach is the use of Markov Random Fields (MRF) to represent the relationships between cells in the RPM grid.

Knowledge Representation

Each cell in the matrix is treated as a node in an undirected graph. The edges between these nodes represent "affinity factors," which are mathematical functions measuring the compatibility of two images based on pixel-level agreement.

Example of two cells (A and B) demonstrating high structural affinity due to identical geometric features.

Pattern Learning and Rule Mapping

Rather than assuming a predefined rule, the system explores various graph structures to identify the "minimal map" that best explains the observed data. This process is guided by principles of parsimony: preferring the smallest group of images and favoring structures where features remain invariant.

These statistical affinities are mapped to classic cognitive rules, such as "constant in a row" or "figure addition." For instance, "Figure Addition" is modeled as a network topology where the third image represents a statistical juxtaposition of the preceding two.

Pattern Recognition

To solve a problem, the model evaluates each potential answer by substituting it into the blank cell and calculating a likelihood score. The candidate that maximizes the logical consistency of the graph—the "identity function"—is selected.

A candidate answer (2) that would yield low likelihood if the model induces a constancy rule, as the 'X' feature breaks the established pattern.

Experimental Evaluation

The model was evaluated using Sets B, C, D, and E of the Standard Raven Progressive Matrices (SPM), totaling 48 problems. Set A was excluded as it primarily tests visual textures rather than the geometric pattern learning the model was designed to address.

Results

The Structural Affinity method correctly solved 44 out of 48 problems (91.67%). Furthermore, it demonstrated high precision in rule induction, achieving a score of 0.94.

Model Accuracy (SPM)
Structural Affinity 91.67% (44/48)
CogSketch 91.67% (44/48)
Fractal 70.00% (42/60)
Affine 65.00% (39/60)
Anthropomorphic 77.78% (28/36)

The model's performance is competitive with CogSketch, a sophisticated model using qualitative spatial representations, and significantly outperforms purely visual models like the Fractal and Affine approaches.

Discussion and Conclusion

The success of the Structural Affinity method suggests that logical deduction can be effectively modeled as statistical inference. By using objective functions to implement human-like heuristics, the model avoids the need for complex linguistic or symbolic engines.

Current limitations include sensitivity to image misalignment and a lack of optimization for complex textures. Future work will focus on extending the model to the Advanced Progressive Matrices (APM) and incorporating color processing for the Colored Progressive Matrices (CPM).

Summary updated

BibTeX
@inproceedings{shegheva2018structural,
  author = {Snejana Shegheva and Ashok K. Goel},
  title = {{The Structural Affinity Method for Solving the Raven's Progressive Matrices Test for Intelligence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11323},
}
Figure