Types of learning that infants might be pretty good at.

OK so, following on from my previous post, here is my attempt to describe five types of learning that infants might be pretty good at. These are based on the distinction that I noted previously between “A high energy barrier [that] favours robust online maintenance of information and a ... low energy barrier [that] is beneficial for flexible and fast switching among representational states”. Babies, hypothetically, are bad at maintaining a high energy barrier state, but might be good at tasks requiring a low energy barrier state.

For example:

1) we might present a variety of different stimuli one by one, across discrete but repeated trials. (The idea for this was actually not mine but John Duncan’s.) The different stimuli vary according to two dimensions – say shape, and colour. Each object then disappears and reappears in either the left location or the right location. Is there a pattern that predicts which object is going to appear where?

In doing this we can contrast three types of patterns:

Condition 3 is the baseline. In this, there is no pattern, and the only way to solve the task is to memorise where each object goes, individually. In condition 1 there is a pattern that can be spotted. If I pay attention just to colour, and ignore shape, and if I remember that green and yellow go left, then I can solve the task with 100% success, and I’ve saved a lot of work, compared to condition 3. So we can predict that condition 1 ought to be performed better than condition 3, and that, between individuals, this difference ought to index how good people are at spotting and applying patterns.

The interesting distinction I think, though is not between condition 1 and condition 3, but between condition 2 and condition 1. In condition 2 there is a pattern, but the required information is spread across multiply covarying stimulus dimensions (just as in a real-world category, such as the difference between cat and dog). The optimal strategy is to juggle your attention between the two dimensions at once. I would predict that babies would be bad at spotting a condition 1 type pattern, and relatively less bad at spotting a condition 2 type pattern. Kloos & Sloutsky have come up with a nice formalisation of this difference, which maps on what some people call 'implicit' learning (condition 2) and 'explicit' learning (condition 1).

Of course, this is not a good experiment in lots of ways – it’s much too hard, the predictions are too complicated, based on multiple subtractions (which is always a noisy design). And it's not quite clean anyway. So...

2) you could do a similar type of task to that described above – but using objects for which it’s not clear where the stimulus dimensions lie. See the stimuli below (these are taken from the same Kloos & Sloutsky paper, although I’m using them in a different context). Do I categorise the objects according to body colour, number of dots, wing length, tail length, number of nodules on the wing, or so on?

When I’m doing a task like this, as an adult I’m aware that I can consider candidate dimensions one by one (i.e. let a few trials go past, evaluating whether body colour might predict the answer), but by the time I’ve rejected that hypothesis I can’t then go back and re-evaluate those trials again, using a different hypothesis). Whereas babies might be able to evaluate more different possible criteria concurrently.

3) we could take a similar idea, but in the time dimension. Play a stream of artificial language (F-A-H-G-D) some have a higher probability of being joined to one another than others do (see e.g.s here). So you have chunks of F-A-H that come up more often than chance, as do chunks like D-A. So when you are listening to the stream you have to go through and subdivide it all through one way and say ‘OK, I got these chunks that appeared to come up quite frequently', and then go through again and parse it another way and say ‘OK but when I parsed it this second way I got more repeated units’. Again, it’s this idea of multiple concurrent revaluations according to different criteria that infants might be good at.

4) Another way of exploring the same type of distinction comes from a paper by Hayes & Broadbent. Here, the task was to respond to outputs from a ‘computer person’ whose ‘anger levels’ went up and down depending on their responses. In the ‘explicit learning’ condition, the output level is set off the participant’s previous response. In the ‘implicit learning’ condition, the output level was set off the participant’s previous response, with a variable time lag in addition. Again, the ‘implicit’ learning condition is what you’d predict that infants might be relatively good at.

5) The fifth type is similar to the distinction often used by Yuko Munakata and others. For example, they found that good performers on dimensional card sorting tasks are faster on delayed match to sample tasks than perseverators because they proactively exercise cognitive control, but under conditions of distraction they are proportionately more slowed. Similarly, another group found that individuals with reduced online goal maintenance abilities showed impaired set-shifting, but that the same individuals also showed reduced deleterious effects of intervening task-irrelevant distractors on a spatial working memory task.

These are all different ideas – and all, in various ways (apart from 3) impractical for use with babies. But they are all potentially, ways of getting at the types of learning that infants might be good at. And formalizing how the differences between the types of tasks at which babies are incredibly bad, and those at which they’re incredibly good, might all lie on one dimension. Babies might be good at 'low energy barrier' learning, that requires juggling between multiple different representations or task structures in order to find the one that 'fits'. (This is the type of learning required, for example, in early stages of language acquisition.) But they might be very bad a 'high energy barrier' learning, that requires paying attention to one thing and ignoring others. (As required, for example, in a working memory task, or in an IQ test.)