Attentional Mechanisms for Learning Feature Combinations

When we make choices in our daily lives, we typically have to consider multiple features of the possible options. Even simple choices, such as choosing a wine to drink, can present us with numerous options that vary in multiple dimensions (e.g., sweetness, region, grape variety, price). However, the outcomes of our choices often depend on only a subset of these features or perhaps specific combinations thereof. A fundamental challenge for any learner is to discern which aspects of a choice or environment are associated with rewarding outcomes (Sutton and Barto, 2018). To cope with complexity, organisms can use simplifying strategies. Instead of learning the value of every single object, we can learn the value of individual features (Niv et al., 2015). For example, rather than tasting every possible bottle of wine, we might focus on just one feature, for example, choosing a sweet rather than dry wine. This drastically reduces the dimensionality of the decision. After trying a few bottles, patterns may emerge, such as “sweet wines are better than dry wines,” allowing the learner to make informed choices even when encountering unfamiliar wines. This feature-based learning strategy facilitates faster decision-making and more efficient generalization to new products. However, feature-based learning itself has limitations (Radulescu et al., 2021). Sometimes, the value of an object is not determined by its individual features but by their specific combinations or conjunctions of features. For example, a person who generally likes sweet wines might discover that they do not like those that come from Bordeaux. They might eventually learn that they like sweet wines from California but dry wines from Bordeaux. In other words, neither “sweetness” nor “region” alone defines a wine they like. The learner instead needs to know the exact conjunction of sweetness … Correspondence should be addressed to Kianté Fernandez at kiante{at}ucla.edu.