Probabilistic classification learning in amnesia.

Amnesic patients and control subjects participated in a study of probabilistic classification learning. In each of three tasks, four different cues were each probabilistically associated with one of two outcomes. On each trial, the cues could appear alone or in combination with other cues and subjects selected the outcome they thought was correct. Feedback was provided after each trial. In each task, the amnesic patients learned gradually to associate the cues with the appropriate outcome at the same rate as control subjects, improving from 50% correct to approximately 65% correct. Presumably because the cue-outcome associations were probabilistic, declarative memory for the outcomes of specific trials was not as useful for performance as the information gradually accrued across trials. Nevertheless, declarative memory does appear to make a contribution to performance when training is extended beyond approximately 50 trials, because with further training control subjects eventually outperformed the amnesic patients. It was also demonstrated that performance on the probabilistic classification task was not the result of holding knowledge of cue-outcome associations in short-term memory, because both control subjects and amnesic patients demonstrated significant savings when testing was interrupted by a 5-min delay (experiment 2). Probabilistic classification learning appears to provide an analog in human subjects for the habit learning tasks that can be acquired normally by animals with hippocampal lesions.