 
| |
|
| |
|
| |
|
| |
The Gluck and Myers model of cortico-hippocampal function can be most easily understood through a concrete example of its application to a form of classical Pavlovian conditioning: motor-reflex conditioning of the rabbit eyeblink response (Figure 4). The stimulus cues in eyeblink conditioning are tones or lights, and the outcome to be predicted from these cues is a mildly aversive airpuff delivered to the rabbits eye. Normally, the airpuff elicits a reflexive blink, protecting the eye from possible damage. When the rabbit learns that a tone predicts an airpuff, it generates a conditioned eyeblink response in anticipation of the airpuff. From observing the strength or probability of this anticipatory eyeblink, we measure the animals association between the tone and the airpuff.
The application of our theory to motor-reflex conditioning can be seen in Figure 5A. The network on the left is the locus of long-term memory storage. It contains the essential pathways for expression of the learned behavior. For most kinds of motor-reflex conditioning, including eyeblink conditioning, this would correspond to the cerebellum, as demonstrated by the work of Richard Thompson and colleagues at the University of Southern California.
Although the long-term memory network on the left of Figure 5A has an internal layer of nodes much like the network in Figure 3, this particular network is unable to develop new representations without a teaching input from the hippocampal network on the right of Figure 5A. This hippocampal region is presumed to be monitoring incoming information from all the senses, including the experimental cues (e.g., tones and lights) as well as the background cues (context). From these sensory inputs, the hippocampal region derives an appropriate representation for each stimulus based on regularities in the stimulus environment. These new representations enhance predictive differentiation and redundancy compression, as described earlier. Information about the appropriate representation of stimuli is then broadcast from the hippocampal network on the right to the long-term memory network on the left, in the form of a teaching signal that tells the long-term memory system how it should represent new stimulus information.
Figure 5B illustrates what remains after the hippocampal inputs to long-term memory are removed. Without any teaching input from the hippocampus, the long-term memory representations are presumed to be fixed. Learning in this system can still occur, but only to the extent that the representations established prior to the hippocampal removal are sufficient. This implies that learning in the lesioned model of Figure 5B will be qualitatively different than learning in the intact model of Figure 5A, because the lesioned learning will not involve representational changes that reflect regularities among the stimuli.
We demonstrate later in this paper that this loss of representational flexibility in the lesioned model does not usually impair the acquisition of simple tone-airpuff associations, although performance on tasks that are more complex, and that do require the development of new and novel representations, may be impaired. In summary, the lesioned model of Figure 5B behaves much like what psychologists from the early part of this century, like Clark Hull, called a stimulus-response system. In contrast, the intact model of Figure 5A combines this stimulus-response learning with a hippocampal-dependent system for learning multiple relationships among all stimuli in the world, akin to Edward Tolmans theories of cognitive maps in the 1940s. To better appreciate the importance of the representational changes that depend on the hippocampal region, we turn now to describe a novel prediction of our theory for a form of transfer generalization between two learning tasks.