From Pigeons to Humans: Grounding Relational Learning in Concrete Examples

Marc T. Tomlinson, Bradley C. LoveAAAI 2006

Introduction

A fundamental question in cognitive science is how organisms transition from perceiving concrete features—such as color or shape—to understanding abstract relations, such as "same," "different," or "larger than." Traditionally, research has been bifurcated: category learning models focus on grouping items by shared features, while analogy models focus on recognizing shared structures. Tomlinson and Love (2010) propose a unified framework called BRIDGES (Building Relations through Instance Driven Gradient Error Shifting), suggesting that abstract reasoning emerges from comparing new experiences to stored concrete examples through a process of structural alignment and attentional shifting.

The BRIDGES Model

BRIDGES extends exemplar-based neural networks, such as the ALCOVE model, by incorporating structural alignment. Unlike models that rely on innate algebraic rules or pre-defined "identity detectors," BRIDGES stores specific instances in memory and evaluates new stimuli based on their similarity to these stored exemplars.

The model operates through three key mechanisms:

  1. Exemplar-Based Representation: Every experienced instance is stored as a distinct node in memory.
  2. Structural Alignment: When encountering a new stimulus, the model performs a mapping to find the optimal one-to-one correspondence between the elements of the new stimulus and those of stored exemplars.
  3. Attentional Shifting: Using gradient descent, the model adjusts attention weights. If featural information (e.g., specific shapes) is unreliable for a task, the model shifts attention toward relational patterns, such as parallel connectivity.
The BRIDGES model shifts attention between featural matches and relational (parallel) connectivity to solve abstract tasks.

Experimental Validation

The authors validated BRIDGES by simulating two landmark behavioral studies that span the spectrum of relational complexity.

Infant Grammar Learning

In simulations of Marcus et al. (1999), the model was exposed to "sentences" of speech sounds following specific patterns (e.g., AAB like "GA TI TI"). BRIDGES successfully replicated infant behavior, showing a preference for novel sounds that followed the learned relational pattern over those that did not. This suggests that "grammar" can be learned through similarity to concrete examples rather than abstract rule-following.

Same-Different Discrimination

BRIDGES was also applied to the same-different discrimination task used with pigeons and humans (Young & Wasserman, 1997). Subjects were trained to distinguish between displays of 16 identical icons ("same") and 16 unique icons ("different").

An example stimulus from the same-different discrimination task (14-1-1), illustrating the graded nature of relational categories.

The model achieved a near-perfect fit ($R^2 = 0.95$), correctly predicting the probability of "different" responses across various intermediate conditions (e.g., displays with 12 identical and 4 unique icons).

Discussion and Key Insights

The success of BRIDGES highlights that abstract categories are often graded rather than absolute. A display with 14 identical icons is perceived as "more same" than one with 8, a nuance captured by the model’s similarity-based approach. Furthermore, the model demonstrates that the transition from featural to relational learning is a matter of shifting attention rather than changing underlying cognitive architecture.

While BRIDGES provides a powerful unification of category learning and analogy, it faces computational challenges. The exhaustive search for optimal mappings is expensive for complex scenes, suggesting a need for heuristic-based pruning in future iterations. Additionally, while the model can learn to attend to relations, it currently requires those relations to be present in the initial representation. Future work will explore the autonomous discovery of relational predicates and the application of these principles to large-scale machine learning tasks.

Summary updated

BibTeX
@inproceedings{tomlinson2006from,
  author = {Marc T. Tomlinson and Bradley C. Love},
  title = {{From Pigeons to Humans: Grounding Relational Learning in Concrete Examples}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure