 
| |
|
| |
|
| |
|
| |
The evidence for predictive differentiation and redundancy compression in associative learning comes from analyses of animal learning in which our theory expects the effects of predictive differentiation or redundancy compression to be most evident. We illustrate the general approach here with reference to one specific task, learned irrelevance. Learned irrelevance and many similar tasks can be understood as resulting from hippocampal-dependent redundancy compression, and should therefore be absent in animals with lesions to the hippocampal region.
TABLE 1. Learned Irrelevance Training
Phase 1
Experiment: T...P..P.T...P..T..T....P...etc
Control: ...(no pretraining).......
Phase 2:
Experiment: TP....TP.....TP...TP.....TP...TP..etc
Control: TP....TP.....TP...TP.....TP...TP..etc
As schematized in Table 1, learned irrelevance involves two phases of learning: In Phase 1, animals receive multiple presentations of a cue (e.g., tone, T) and an outcome (e.g., airpuff, P), which are presented in an uncorrelated manner, such that the presence of the cue is not predictive of the outcome. In a second phase of training the cue becomes predictive of the outcome so that tone presentations are always followed by an airpuff. Learned irrelevance refers to the finding, in normal animals, that pretraining with the two uncorrelated cues in Phase 1 slows down later learning of the TP association in Phase 2, as compared to control animals who received no pretraining.
Within our theory of hippocampal function, learned irrelevance occurs in the intact model of Figure 5A, because the tone stimulus is redundant with the collection of static cues within the training environment, cues which we collectively refer to as the background context of the experiment. Because they dont change, they provide no additional information for predicting the airpuff. Therefore, by the principle of redundancy compression, we expect the representation of the tone within the training context to become more similar to the representation of the background context alone. This will slow down learning in Phase 2 because of increased difficulty in discriminating between the tone and the context alone. In intuitive terms, the hippocampal-dependent process of redundancy compression is causing the irrelevant tone to disappear into the background context because it is not usefully informative.
A simulation of the intact model in Figure 6A shows that it is slower to learn the toneairpuff association following an uncorrelated pre-exposure (dashed green line) phase as compared to a control condition (solid green line) that was not pre-exposed to the uncorrelated cue and outcome. We can compare the simulated learned irrelevance in our intact model to data collected in our lab in collaboration with M. Todd Allen. Figure 6B shows that intact control animals exhibit the normal learned irrelevance effect.
We next describe a similar study of a human analog of learned irrelevance in normal control subjects, as well as in amnesic patients who have sustained damage to the hippocampus and other structures in the medial temporal lobe of their brain. Later, we return to the animal paradigm to compare the amnesic data to the behavior of rabbits who have undergone surgical removal of their hippocampal region.