Many claims about concept learning in animals rely on binary categorization tasks (Herrnstein et al., 1976; Freedman et al., 2001; Marsh and MacDonald, 2008). When subjects exceed chance levels of performance, they are alleged to have learned “the concept.” Critics are quick to point out that although subjects have learned something, confounds may explain performance more simply (Katz et al., 2007; Wright and Lickteig, 2010; Zentall et al., 2014). Despite a growing literature on both sides, supporters of “concept learning in animals” seem no closer to persuading the skeptics, while skeptics are no closer to persuading proponents. This rift hinges on disagreements over the strength of the evidence. Results from dichotomous classification procedures represent the weakest possible evidence for concepts in animals, for reasons unrelated to the validity of corresponding theoretical claims. One pitfall is the tailor-made classifier, which may arise during training. Effectively, “teaching to the test” undermines claims about animals' general knowledge. Another is the lucky guess during testing. A simplistic response during the testing phase will yield many rewards due to guessing alone, making it difficult to assess the precise content of learning. These shortcomings are independent, such that either might confound an experiment. 1.1. The tailor-made classifier The risk of animal subjects “outsmarting” their minders has been with us since Clever Hans. Whatever the aims of our experimental paradigms, the influence of extraneous information must be minimized so that results reflect the intended empirical test. Concept learning presents the scrupulous researcher with a challenge: How does one identify (much less control for) the extraneous features of a stimulus? Our understanding of how the brain categorizes stimuli remains limited (Freedman and Assad, 2011), but there is also no consensus about what constitutes a feature. The list of stimulus attributes that might be used to categorize stimuli includes overall descriptive statistics (“presence of the color green”), low-level structural details (“T-shaped edge junctions”), patterning (“presence/absence of tiled features”), functional interpretation (“looks like food”), ecological indicators (“bright color = poison”), and interacting levels of analysis (cf. Spalding and Ross, 2000; Marsh and MacDonald, 2008). As such, the content of learning is subject to multiple interpretations. A classifier is an algorithm (however simple or complex) that matches a stimulus with a discrete category. In general, classifiers must undergo training to become sensitive to category-relevant features. Any classifier is limited in what it can detect, and some begin with innate knowledge (such as instincts that some stimuli are threatening). These characteristics hold both for computer algorithms and for the processes used by organisms to classify stimuli. The aim of studying how organisms solve problems of this kind is to discover and describe their classifiers, and to distinguish processes that have evolved recently from those that are well-preserved across many species. Herein lies the problem: When training requires that only two categories be identified, then the classifier (and therefore the organism) need only identify some difference that distinguishes them, and nothing more. The result is a tailor-made classifier: Tailored by the specifics of the binary training paradigm, and optimized solely for that dichotomous discrimination. Just as a bespoke suit is tailored upon request to fit a single person, a tailor-made classifier is only effective at the discrimination it was trained for. At its most extreme, this is Clever Hans in a nutshell: A (cognitively) cheap trick that yields rewards but falls short of generalized knowledge. When faced with this problem, researchers often narrow the scope of the features available. A set of images might have colors removed, luminances matched, occluders introduced, and noise added (e.g., Basile and Hampton, 2013). Such studies are valuable because they help reveal which features can be used by the classifier. However, regularized stimuli cannot rule a tailor-made classifier, because so many potential “features” might provide the basis for the classification. Furthermore, insofar as the resulting stimuli are “unnatural,” they generalize poorly to how stimuli are categorized in ecological contexts. So long as any feature is consistent enough across stimuli to permit classification, there is a possibility that the classifier relies exclusively on that feature. To be sure that the classifier used by an organism is capable of generalized knowledge, it is essential that training encompass more than a single dichotomous categorization.
Read full abstract
Sep 20, 2024
Sandeep Kumar + 1
Open Access