Learning how to learn: Learning sets

The concept of learning set that was suggested by Harlow (1949) more than a half a century ago, still bridges a gap between the behaviourists’ theory of learning by trial and error and the Gestalt hypothesis that learning could be achieved suddenly, or insightfully. Harlow hypothesised that trial-and-error learning and insight learning are related and that insight develops out of well established connections between stimuli and reactions (S-R connections). He suggested the concept of learning set to refer to an intervening period between S-R and insight. In other words, as Harlow gives it, when an animal learns a new kind of problem, it solves it by slow painful plodding trial and error. However, if it had experience with a large number of problems of a single type then problems of this type can be solved insightfully. These abilities enable animals to adapt quickly to new environments, given that the individual is already equipped with a set pattern of understanding situations and solving problems in principle. This theory, " Learning to Learn", described the ability of animals to slowly learn a general rule that could then be applied to rapidly solve new problem sets.

Harlow suggested that learning set formation could serve as a relevant measure of intelligence, that is, individual’s ability to solve complex problems by using representations of previously experienced events. He claimed that performance levels on DLS (Discriminative Learning Set) tasks across species are generally in keeping with an ordering base on cortical complexity (Harlow, 1959). Modern studies in cognitive ethology have demonstrated that the use of a single measure such as learning set formation is unlikely to test adequately all aspects of animal intelligence (Pearce, 2000). Nevertheless, an experimental paradigm elaborated for exploring learning set formation, is widely accepted in cognitive studies.

Harlow used WGTA apparatus (see Part III) to establish experimental procedure for discrimination. This allowed asking the question of how quickly and efficiently animals can learn to evaluate hypotheses. Harlow conducted his experiment so that his animal subjects (typically monkeys, but the same results were found in a number of other species) had to learn a long series of simultaneous discrimination problems. The monkey was first presented with two stimuli (a red block and a thimble, for example); one was predetermined " correct" and reinforced with food (red block) and the other was " incorrect" and not reinforced with food (thimble). After each selection, the objects were replaced and the monkey again chose a stimulus. Each trial reinforced the same stimulus (red block). The monkey had a 50% chance of being " correct" on each trial; however, it could increase its chances by adopting the win-stay, lose-shift strategy. That strategy is in some sense the backbone of a hypothesis testing approach: As long as what you are doing is correct, stick with it (win-stay); when it leads to a wrong answer, change it (lose-shift). For example, if the monkey chose the thimble and was not reinforced, it should shift to the red block for the award. If, however, it correctly selected the red block and was reinforced, it should stay with the reinforced stimulus and choose the same stimulus next time. The monkey continued throughout a series of six trials with eight pairs of stimuli (learning sets). Harlow found the monkeys to be averaging approximately 75% correct responses by the sixth trial of the eighth set. He then began to look at the animal's behaviour during the second trial with the new stimuli, say, a box of matches and a half of tennis ball. He found the monkeys to implement the stay or shift strategy on the second trial of the six-trial set, which means the animals did not relearn the strategy with each new set of stimuli, they instead applied the rule they had already learned. After 250-plus trials, the monkeys were about 98% correct on the second through the sixth trials each with a new stimuli set.

Harlow also asked whether his animals could acquire a win-shift lose-stay strategy. In this latter series of studies, a problem is always followed by a reversal shift problem with the same stimuli, after which a new problem is given, and then followed by its reversal shift. So, once you have solved a problem with new objects, you will now be placed on the next problem in which you will need to choose the stimulus exactly opposite to the one you just picked. In a particularly devious version of this procedure, we get you to the point where there are just two trials on each set of stimuli, the second being the reversal shift. Here, to do well, you must immediately pick the other stimulus if you choose correctly on the first trial (win-shift), and continue to pick the same stimulus if you chose incorrectly (lose-stay), since that loser will be the winner on the next round. Studies based on reversal learning have shown that apes, macaques and capuchin monkeys can generate abstract rules of learning such as “select the object that was previously incorrect” (Rumbaugh and Pate, 1984; Washburn and Rumbaugh, 1989; DeLillo and Visalberghi, 1994).

Examining animals’ ability to establish learning sets can be considered an essential part of testing their ability to create abstract concepts and, in a general sense, to grasp regularities. Experimental evidences of these abilities in different species have been described in Chapter 14. For example, in the experiments of Sappington and Goldman (1994) on discrimination learning in Arabian horses) cited above, the first – very simple – discrimination took the longest to learn, and was followed by a general decrease in the time to learn successive problems. This suggests that the subjects were able to use previously learned information to facilitate subsequent learning, and supports the conclusion that horses have the ability to “learn to learn”. Horses do this in the domestic training environment when they master advanced manoeuvres faster after learning a series of preliminary tasks. There are some experimental evidences of learning to learn in horses tested on serial reversals of a positional discrimination (Warren, 1965). The subjects quickly exhibited a rapid decline in error rates on consecutive reversals.

It is of no surprise that practically all students of cognitive skills in animals note that their subjects display learning set formation with the course of multi-stage experiments. Even Skinner’s followers regard the phenomenon of learning to learn when shaping animals’ behaviour in accordance with principles of operant conditioning. For instance, in experiments on shaping new (unnatural, novel, untrained) behaviour it took only a few days of continuously repeating the first step in conditioning (reinforcing new or more specific behaviours) for dolphins to emit an unprecedented range of behaviours (Pryor et al., 1969; Herman, 1980; see details in Chapter 6). By a few number of sessions, it had become apparent that dolphins not only knew that only novel behaviour would be reinforced, but also that only one behaviour per session would be reinforced. Dolphins and other species easily applied the gained experience to grasp rules for other games with their trainers that included other requests for animals. These laboratory studies reflect many situations in nature where gained experiences facilitate animal’s ability to solve new problems.