DISCRIMINATION LEARNING
In a discrimination learning task the animal is presented with two stimuli (sometimes more) that are associated with different outcomes. For example, a pigeon might be presented with a choice between two discs, one coloured red and the other green; pecking at the green disc will produce food, but pecking at the red disc will not. The pigeon will solve this problem, coming reliably to choose the green disc after a few dozen trials. Its ability to do this task is no puzzle and can be fully explained in terms of standard conditioning processes. More intriguing is the fact that training on such a task will transfer to other similar tasks. If the pigeon is now asked to solve a similar discrimination problem, in which the choice is between blue and yellow discs, learning can occur very rapidly: we call this positive transfer. The original associations involving red and green are clearly irrelevant to this new discrimination task, so the transfer must have some other source. The pigeon appears to have acquired a fairly abstract concept in the course of acquiring the first discrimination – something along the lines of ‘differences in colour are important and should be attended to’. Studies involving primates have produced more dramatic examples of abstraction. In the learning-set procedure (first introduced by Harlow, 1949), a rhesus monkey is presented with two objects and given a small amount of food for picking up one of them. After six trials the original objects are replaced with two new ones and, again, responding to only one of the objects was rewarded. After six trials on this new problem, the objects were again changed, and so on for many, many pairs of objects. Early in training performance is unremarkable, six trials being insufficient for the monkey to solve the problem. But as training proceeds, performance begins to improve, until finally it is as near perfect as it can be (see figure 4.14). After training on hundreds of these problems, the monkey is able to solve a new problem with no more than a single error, switching its choice to the other object if its first choice is wrong, but staying with its original choice if this proves correct. By experiencing many problems of a similar type, the animal appears to abstract some general rule about how to behave in this situation – a rule that allows the near-instantaneous solution of a problem that it had, in fact, never faced before. The rule that operates in this case is the win-stay, lose-shift strategy: in other words, the animal learns to persist with a choice that yields food, but shift to the other object if it does not. Associative theory can go some way towards explaining this. The occurrence of reward (or non-reward) can be regarded as a stimulus that, like any other, can enter into associations or acquire discriminative control over an instrumental action. The special feature of the learning-set procedure is that these stimuli and associations come to dominate the animal’s behaviour to the exclusion of all others. So the animal learns to focus on classes of cues that are accurate predictors of reward and to ignore others that are not. Intensive research is currently going into the nature of such higher-level learning processes that might modulate the mechanims of simpler associative processes.
In a discrimination learning task the animal is presented with two stimuli (sometimes more) that are associated with different outcomes. For example, a pigeon might be presented with a choice between two discs, one coloured red and the other green; pecking at the green disc will produce food, but pecking at the red disc will not. The pigeon will solve this problem, coming reliably to choose the green disc after a few dozen trials. Its ability to do this task is no puzzle and can be fully explained in terms of standard conditioning processes. More intriguing is the fact that training on such a task will transfer to other similar tasks. If the pigeon is now asked to solve a similar discrimination problem, in which the choice is between blue and yellow discs, learning can occur very rapidly: we call this positive transfer. The original associations involving red and green are clearly irrelevant to this new discrimination task, so the transfer must have some other source. The pigeon appears to have acquired a fairly abstract concept in the course of acquiring the first discrimination – something along the lines of ‘differences in colour are important and should be attended to’. Studies involving primates have produced more dramatic examples of abstraction. In the learning-set procedure (first introduced by Harlow, 1949), a rhesus monkey is presented with two objects and given a small amount of food for picking up one of them. After six trials the original objects are replaced with two new ones and, again, responding to only one of the objects was rewarded. After six trials on this new problem, the objects were again changed, and so on for many, many pairs of objects. Early in training performance is unremarkable, six trials being insufficient for the monkey to solve the problem. But as training proceeds, performance begins to improve, until finally it is as near perfect as it can be (see figure 4.14). After training on hundreds of these problems, the monkey is able to solve a new problem with no more than a single error, switching its choice to the other object if its first choice is wrong, but staying with its original choice if this proves correct. By experiencing many problems of a similar type, the animal appears to abstract some general rule about how to behave in this situation – a rule that allows the near-instantaneous solution of a problem that it had, in fact, never faced before. The rule that operates in this case is the win-stay, lose-shift strategy: in other words, the animal learns to persist with a choice that yields food, but shift to the other object if it does not. Associative theory can go some way towards explaining this. The occurrence of reward (or non-reward) can be regarded as a stimulus that, like any other, can enter into associations or acquire discriminative control over an instrumental action. The special feature of the learning-set procedure is that these stimuli and associations come to dominate the animal’s behaviour to the exclusion of all others. So the animal learns to focus on classes of cues that are accurate predictors of reward and to ignore others that are not. Intensive research is currently going into the nature of such higher-level learning processes that might modulate the mechanims of simpler associative processes.
No comments:
Post a Comment