Feature attention during learning and generalization

The world is rich with more information than people are capable of processing. By selectively attending to features that organize causal relationships, and by modulating attention according to immediate goals, people learn useful and generalizable internal representations.
Feature attention during learning and generalization

No one is born with extensive knowledge of chocolate cereal. Instead, people form internal representations of chocolate cereal through taking actions (pouring in milk versus mixing with yogurt) and experiencing outcomes (delicious versus disturbing).

The features people attend to when identifying a chocolate cereal depends both on consistency across past examples (brown color), and the feature’s informativeness for the goal of distinguishing between alternatives in a specific context (features that distinguish chocolate cereal from chocolate crackers). This leads to several questions (see Figure 1):

  • What cognitive processes form internal representations of chocolate cereal? 
  • How do people ignore irrelevant features when generalizing to new brands of chocolate cereal? 
  • The first time a person has to determine if an item is a chocolate cereal or chocolate crackers, how do they narrow their attention to the features that help them avoid pouring milk on the crackers?

We developed a computational algorithm that learns to represent the environment by taking actions and experiencing outcomes. In more technical terms, it is a latent-state instrumental reinforcement learning model that balances creation of new internal representations with generalization of existing internal representations (terminology defined in the Table).




The set of potential options available during a decision. For example, items in the breakfast aisle of a grocery store.


The consistency of features across past examples. If all chocolate cereals are encountered in boxes of similar sizes, then the box and the size have high covariance. 


An individual example generated by a latent state, such as the specific bowl of chocolate cereal on the table. 

Goal-directed attention

Attention derived from immediate goals in a context, rather than bottom-up stimulus properties. When the goal is determining if a cereal is chocolate or wheat, the fact it is in a bowl is irrelevant; however, when discriminating between chocolate cereal or a collection of small brown rocks, whether it is in a bowl (versus on the ground) is essential.

Instrumental reinforcement learning

A model of learning where values are updated through taking actions and experiencing outcomes. For example, pouring milk on chocolate cereal and tasting something delicious, versus mixing it with yogurt and reaching for the trash can. 

Internal representation

How latent-states of the world (e.g., chocolate cereal) are represented within an animal or computational algorithm.

Latent state

A state that cannot be directly observed (i.e. latent), but is inferred through experiencing its causal statistics. For example, the collection of unobserved properties that give rise to what we call chocolate cereal. 


An internal representation formed through summary statistics, such as the covariance of features. A chocolate cereal prototype would involve mean properties such as brown color and the general size.

The model combines attention to features common across past encounters with attention to features informative for immediate goals (see Figure 2).

We then tested model predictions through three experiments with human subjects. 

The first experiment tested subjects' use of goal-directed features. The task was similar to generalizing an internal representation of chocolate cereal to a brand of chocolate cereal never before encountered. We found that most subjects were able to generalize. Moreover, the gradient in generalization ability we observed could be accounted for by manipulating goal-directed attention in our computational model. 

The second and third experiments tested whether subjects exclusively attended to the goal-directed features, whether they memorized each example as distinct internal representations, or whether they formed internal representations (prototype/exemplar) consistent with the latent-states in the environment. The “context generalization” task used in these experiments are similar to the case where one learns to tell chocolate cereal from wheat cereal in the breakfast aisle, and chocolate crackers from wheat crackers in the snacks aisle. When an item is then found on a friend’s kitchen counter (a novel context), one must determine if it is chocolate cereal or chocolate crackers.

Experiments two and three found distinct subpopulations of subjects

One subpopulation primarily attended to the features that were goal-directed during initial learning. Behavior of these goal-directed attention subjects was largely reproduced through manipulating the context recognition mechanism in the model. 

A second subpopulation attended to the covariance and mean of features within each state. Their behavior required a model with prototype state representations and errors based on state estimation, rather than action selection. 

A third subpopulation attended to feature covariance, but made errors that were largely random. Their behavior also required a prototype model. 

Thus, these experiments found that while subjects learn states defined by the covariance of features, there is diversity in their use of goal-directed attention, and in the structure of their errors (see Figure 3).

Taken together, our findings suggest people learn to attend to prototype features, and have strong inter-individual differences in goal-directed attention. Furthermore, we provide a set of novel computational models and associated behavioral tasks. These findings are highly relevant for psychology, neuroscience, machine learning, and anyone partial to crackers or breakfast cereal.

Please sign in or register for FREE

If you are a registered user on Behavioural and Social Sciences at Nature Portfolio, please sign in