Brain Plasticity Through the Life Span: Learning to Learn and Action Video Games - Bavelier et al. - 2012 - Article

Humans are excellent learners and through training they can acquire new skills and alter existing behaviors. However, research shows that learning that emerges through training often does not transfer to other contexts and tasks. Researchers then started focusing on which conditions were necessary to stimulate general learning. These trainings are usually more complex and more similar to real-life situations. Recent research shows that videogame training might be promising in general cognitive training.

Videogame training

Recent research has shown a lot of improvements after videogame training. One reason as to why videogame training seems to promote general learning is that playing videogames incorporates a lot of different tasks and domains, that in laboratory trainings have been separated. Playing action videogames, primarily first person shooters, enhances the spatial and temporal resolution of vision, as well as its sensitivity. Other improvements were in visual short-term memory, spatial cognition, multitasking, some aspects of executive function, reaction time, speed-accuracy trade-off, selective attention, divided attention and sustained attention.

A causal relationship

Some fear that the relationship between videogame training and improvements in cognitive control are actually caused by a population bias. This means that action videogaming tends to attract people with inherently superior skills in those games. The only way to confirm that this relationship is causal, is through well-controlled training studies. In this type of training they take participants who don’t usually play action videogames. They pretest them. Then they randomly select half and let them train in the action videogame. The other half has to play a game as well, but it is a nonaction videogame. They compared these two groups and also test for test-retest effects.

Learning to learn as the common cause

A wide range of tasks seem to improve after training in videogames. Researchers wonder what it is in action videogames that improves performance. According to these authors the common cause is learning to learn. First of all, all the trainings share the same principle, namely that the participants need to make a decision based on a limited amount of noisy data. This taps with most everyday decisions.

Posterior distribution over choices

The posterior distribution is the probability distribution of an unknown quantity, treated as a random variable, conditional on the evidence that has been obtained from an experiment or a survey.

In the posterior distribution over choices, we denote the probability as p(c|e). Here c are the choices, and e is the evidence. Now, the most accurate posterior distribution needs to be calculated, so that the best decision can be made. The main goal of learning is to improve the precision of this probabilistic inference.

Research has shown that game experience led to more accurate knowledge of the statistics of the evidence for the task (or more accurate knowledge of the posterior distribution, p(e|c)). Also, most of the tasks that videogame training enhances can be formalized as instances of probabilistic inference. That includes attentional, cognitive and perceptual tasks.

Resources and knowledge

Players of videogames have increased attentional resources in multiple-object-tracking tasks. This allows for more accurate representation of motion and features, which leads to more accurate tracking and identification. Having more resources may enable the learners to learn faster, because critical distinctions will be more accessible to them.

However, only having more resources is not enough to ensure faster learning. Because the resource allocation needs to be guided through structured knowledge, to help find out where the useful information lies. When speaking of knowledge, they refer to it as the representational structure that is used to guide behavior.

Hierarchical behavior models

Hierarchical behavior models divide tasks into subtasks, which are themselves decomposed into component actions. There hierarchical structures allow for the decomposition of computations into multiple layers, using greater and greater abstraction. Shallow architectures abstain from abstraction and simply focus on finding the right set of diagnostic features. This distinction between the shallow and deep hierarchies can be expressed as the difference between learning a rich generative model that captures hidden structure in the data, versus learning a discriminative model specific to a classification problem.

For action videogames to provide the players with knowledge that they can use in their laboratory task, the games and the laboratory tasks need to share structure at some level of abstraction. Otherwise it won’t have any effect on performance in real-life tasks.

Learning rules

There is some evidence pointing to changes in knowledge or learning rules because of playing videogames. First of all it seems that playing action videogames leads to more accurate probabilistic inference. This suggests the development of new connectivity and knowledge that enable a more efficient hierarchy for the task. Secondly there are a few experiments that show improvement despite little need for resources. Thirdly, an attentional explanation is not always in line with the noted changes.

How does playing action videogames lead to general learning?

A possible explanation is that playing action videogames may enable more generalizable knowledge through various abstractions. This includes the extent to which nontask-relevant information should be suppressed, how the performance needs to be modified to maximize the reward rate, how the data needs to be combines across feature dimensions, and how to set a proper learning rate.

According to the authors, playing action videogames does not teach any particular skill on its own. Instead, it increases the ability to extract patterns or regularities in the environment. Players of action videogames have a higher ability to exploit task-relevant information more efficiently. They are also better in suppressing irrelevant information. They might be better at this because they are better in finding out the underlying structure of the task they need to perform. Because they have more accurate statistical inference over the data that they are experiencing, they perform better on a variety of tasks. This is how playing action videogames stimulates learning to learn.