Abstract
We examine online learning in the context of the Wisconsin Card Sorting Task (WCST), a task for which the concept acquisition strategies for human and other primates are well documented. We describe a new WCST experiment in rhesus monkeys, comparing the monkeys' behaviors to that of online learning algorithms. Our expectation is that insights gained from this work and future research can lead to improved artificial learning systems.
Introduction
In standard machine learning, data is often assumed to be stationary. However, real-world environments frequently exhibit concept drift, where the underlying rules governing correct behavior change over time. The Wisconsin Card Sorting Task (WCST) is a classic psychological tool used to measure "set shifting"—the ability to adapt to such changes. This study bridges primate cognition and computational learning theory by modeling monkey behavior in the WCST as a formal online learning problem.
Experimental Methods
The study involved seven rhesus macaques (Macaca mulatta) performing a simplified WCST on a touchscreen. In each trial, monkeys were presented with three objects, each defined by a unique color (Red, Green, Blue) and shape (Circle, Star, Triangle). The objective was to identify a hidden target concept (e.g., "Red"). Once a monkey achieved ten consecutive correct responses, the target concept shifted (e.g., to "Triangle") without explicit signaling.
The task was modeled as a disjunction learning problem using 6-dimensional Boolean vectors. The researchers compared monkey performance against a hypothesis-elimination algorithm. To better simulate biological behavior, they introduced two parameters:
- Slowness (α): The probability of skipping a learning update.
- Stubbornness (β): The probability of retaining a hypothesis despite contradictory feedback.
Results and Analysis
Monkeys demonstrated significant "inefficiency" compared to optimal algorithms. While an optimal algorithm would make approximately 13 errors across the four concepts, monkeys averaged 732 errors. However, their performance remained far superior to random guessing, which would statistically require over 80,000 trials to achieve the same success criteria.
A critical finding was the prevalence of perseverative errors—the tendency to continue applying a previously successful rule after a concept shift. These errors accounted for roughly 75% of monkey mistakes. By tuning the "slow and stubborn" algorithm parameters (α = 0.93, β = 0.96), the researchers were able to closely replicate both the error rates and the error types observed in the primates, with the model producing 67% perseverative errors.
Discussion
The disparity between monkey and algorithmic performance highlights a "game knowledge gap." While the algorithm is pre-programmed with the hypothesis space (knowing the rule is one of six features), monkeys must learn the relevant dimensions through experience. The success of the "stubborn" model suggests that what appears to be biological inefficiency may be a structural component of primate learning, potentially offering robustness in noisy environments where sensor noise might otherwise cause an algorithm to abandon a correct hypothesis too quickly.
Future research will focus on how primates acquire the hypothesis space itself and why they do not show significant "learning to learn" improvements across successive concept shifts in this specific experimental design.
Summary updated