« PreviousContinue »
Let me pursue this question with respect to the work on immediate memory. To understand this area you must know some background. The behaviorist era in psychology, which reigned in its various forms for the thirty years prior to World War II, moved the question of learning to be the central question of an objective psychology. The study of sensation and perception gradually came to tale subordinate places. Even more so, the study of memory became simply an aspect of learning. When work on immediate memory was restimulated in the fifties and sixties, it was largely as a re-emphasis within the notion of learning. Thus, these studies could be conducted with only the issues of memory in mind the nature of acquisition, retrieval, capacity, reliability, etc.
If I were to suggest to this audience that they study the structure of an unknown information processing system, then certainly the kinds of memories would be of prime importance, i.e., their capacities and access characteristics. But the nature of the rest of the central processor would be of equal interest, i.e., the control structure and the basic processing operations. Almost none of this concern with processing, as opposed to memory, is evident in the earlier psychological literature on immediate memory. But recently within the sixties there has been a shift toward such concern. And this shift carries with it the use of information processing theories in detail.
Some brief examples are appropriate to show this situation. I will not attempt any historical comparison, but rather give examples of current work that uses information processing assumptions, not as metaphor but as a theory.
If we ask a subject "What is the 7th letter after G in the alphabet?" (Answer: N), it will take him about a second and a half to respond. If we vary this question by changing the starting letter and the number, then we get a curve, such as that shown in Figure 3 for subject RS. If we kept at our subject long enough, we might expect him to memorize all the answers (there are only 26x25 = 650 distinct questions), in which case the time to respond might be independent of the details of the question. But barring that, the subject must somehow generate the The figure immediately suggests that he does this by counting down the
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
FIGURE 3: Average reaction time to count down alphabet (adapted from 01shavsky, 1965, Fig. 2).
alphabet at a constant rate (he is silent during the interval between question and answer, so may proceed any way he wants). That is, we model our subject as a simple serial processing system which has operations of "get next," "add 1," "test if tally = n" and "speak result," along with some control structure for integrating the performance into a repetitive loop. The linearity arises because the same operations are being performed repetitively.
This particular figure, taken from a Masters thesis at CMU (01shavsky, 1965), is not an isolated example. It shows several things that characterize much of the experimental work on the immediate processor. First, the task is very simple, thus illustrating the earlier point that information processing systems, not artificial intelligence systems should be our main concern. Second, the response measure is reaction time, so that the task is to infer the structure of a complex process from the time it takes to perform it. Third, a population of tasks is used, so that some gross aspect, such as the linearity in Figure 3, contains the essential induction from data to mechanism. Since, in fact, reaction times are highly variable, it is this last feature (initiated by Neisser, 1963) which distinguishes current work from a long history of earlier work on reaction times that didn't bear such fruit.
Figure 4, from a study by Sternberg (1967), reinforces these points. He gave his subject a set of digits, say (1, 3, 7), and then asked them if a specified digit, say 8, was a member of the set. He finds, as the figure shows, that not only does it take longer to answer the question for larger sets, but the relationship is linear. Thus, again, the natural interpretation is according to a processing system engaged in repetitive search. (Though the search here is through immediate memory, whereas it was through long term memory in Figure 3.) Now the point of showing this second example is that Sternberg goes on to use this basic result in an ingenious way. In one condition he presents the subject with a fuzzy, degraded image. What should happen?
We know, independently, that it takes longer to compare a degraded image than a clear one to a known digit. One possibility is that the subject works with the
image, thus having to make the more difficult comparison at each step of the search. If this were the case, the slope of the data line should be greater for the fuzzy image than for the clear image. A second possibility is that the subject initially identifies which digit the fuzzy image represents and then compares an internal representation on each stage of the search. In this case, the slope should be the same, but there should be extra time for initialization. As Figure 4 shows, the latter clearly prevails. Thus we can infer that the operation of perceptual identification occurs prior to the search in immediate memory.
The point of this study, for us, is to see how definitely Sterberg is working with a processing model. The situation is so simple that the key properties can be inferred without creating a program to simulate the subject. But the dependence on the detailed theory is no less for that.
I will present you one more example, since I really wish to convince you of the extent to which information processing theories are taking hold at the level of studying the immediate processor. This is work done by Donald Dansereau in a Ph.D. thesis just completed at Carnegie-Mellon (Dansereau, 1969). He studied the process of mental multiplication, e.g., "Multiply 27 by 132 in your head and when you are through, give the answer." His subjects were all highly practiced; even so, it takes a substantial length of time--e.g., about 130 seconds for 27x132. Again, as with these other studies, time was the measure, and he gave his subjects a large population of tasks.
Now the fundamental fact about mental multiplication is that any crude processing model does quite well. That is, a reasonable count of the steps required by the method the subject uses (e.g., 62x943 requires 5 holds for the given digits, 6 single-digit multiplications, 9 additions, 4 carries and 11 holds for a total difficulty factor of 35) does quite well in predicting the time taken. Figure 5 shows actual times taken versus this difficulty factor for a particular subject. The linear regression accounts for about 90% of the variance. However, this result is not at all sensitive to the exact assumptions. Other work has gotten similar results with quite different measures (Thomas, 1963), though in all cases they are crude processing models.