Kevin at IQ’s Corner has blogged about a new paper in PNAS showing that “working memory” training can improve measures of fluid intelligence – a capacity long thought to be relatively insensitive to experience, and intricately tied to the most complex human cognitions like reasoning, planning, and abstraction in novel contexts.
Jaeggi et al., posit that no empirical evidence shows “computer games enhance anything beyond task-specific performance and selective visuospatial attention” (which must bother our friends at Lumosity & SharpBrains – sigh), and highlight the concern that by “practicing to the test” of fluid intelligence, one might reduce the novelty of those tests and thereby also reduce their power to measure fluid intelligence.
Nonetheless, Jaeggi et al endorse a view that “far transfer” – in which the benefits of training can be manifest in tasks which were not practiced or bear little apparent similarity to the trained tasks – is theoretically possible, and perhaps even probable in the case of the most complex cognitive functions which are “capacity limited.” That is, limitations in attention may be the causal path through which working memory capacity and fluid intelligence are linked, since attention is critically required for the “temporary binding processes” which are used to associate particular items with a particular context. Special emphasis is given to the idea that this temporary binding is in fact “temporal binding”: items must be associated with a particular moment in time for the optimal use of capacity-limited resources like attention.
To test this idea, Jaeggi et al trained subjects on a challenging working memory task in 4 separate experiments. The task involved the serial presentation of both visual stimuli (squares in one of eight spatial locations) and auditory stimuli (one of eight possible letters) at a rate of one item every 3s. Subjects were required to identify whether the current stimulus matched the stimulus presented n trials ago in either the visual or auditory modality – known as an “n-back” task. Critically, the value of n was adaptively determined by individual subject performance, such that if subjects made fewer than 3 mistakes for a given n on a block of 20+n trials, n was incremented by 1 for the following block; this culminated into 20 blocks of trials, requiring about 25 minutes of training per session. The experiments differed in the number of sessions provided (anywhere from 8 to 15) but performance was always compared against a control group, who took the same pre- and posttests assessing fluid intelligence (via Raven’s Advanced Progressive Matrices or BOMAT [a harder version of Raven's]).
This training task involves a variety of demands on higher-level cognitive processes: “dual-tasking” in terms of divided attention between modalities, commonly thought to tax so-called “executive” processing; a reduction in task- or stimulus-specific processing due to the variability in n and the use of varied stimuli; the rapid binding of particular items to their temporal order; and the maintenance, manipulation, and monitoring requirements for remembering a list of items, iteratively updating that list with the presentation of new items, and the iterative comparison of new stimuli with those in memory, respectively.
The results showed that the trained group showed larger gains in fluid intelligence than the control group (partial eta squared of .07; even the control group showed some improvement, probably via practice effects). Using pre-test scores as a covariate (an important thing to do, since subjects may differ in their gains based on their pre-existing abilities), training had statistically significant effects after 17 days. In fact, subjects with lower pre-test fluid intelligence scores showed greater gains on the post-test (this was not specific to the trained group, but could reflect regression to the mean).
While the magnitude of training effects did not interact with pre-test measures of working memory (assessed via digit span and reading span), training did improve digit span scores. However, the training-induced enhancement of fluid intelligence (and it’s relation to the number of sessions) was not fully explained by these gains, nor by the maximum n reached in training, as confirmed through analyses of covariance.
In summary, the authors advocate the view that transfer of training to fluid intelligence occurred through the enhancement of controlled attention, as opposed to the more specific cognitive processes involved in their dual n-back task.
One critical note about this and most other training studies showing far transfer: they tend to implement relatively massive manipulations of executive demands. It is not clear whether such breadth of executive demands is required for training to show transfer, or whether some particular aspect of these demands is sufficient for enhancing intelligence. This methodological shortcoming is most likely due to the large amount of time that needs to be invested in a training study; more targeted manipulations might work, but might show a null effect (publish or perish!). I’ll continue to review studies of training working memory and intelligence throughout the week, with an eye towards identifying the necessary and sufficient characteristics for a training regime to show transfer to intelligence tests.
Updating Training Shows only Near Transfer (At the Group-Level)
Dramatic Play and Executive Function Training (with Far-Transfer)
New Approaches to Training Working Memory [A presentation I gave to LearningRX]
Pharmacological Enhancement of Working Memory
Enhancing Memory with Visual Flicker
Book Review: The Future of the Brain
Caffeine: A User’s Guide to Getting Optimally Wired
Filtering Perception to Save Memory
Cytoskeletal Enhancement of Memory
“Deprogramming” Through Meditation and Hypnosis
Traumatic Brain Injury: Interventions and Treatment