Page images
PDF
EPUB
[merged small][merged small][merged small][ocr errors][merged small][merged small][merged small][merged small][merged small][ocr errors][merged small][ocr errors][ocr errors][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small]

FIGURE 5: Memory transfer rates in mental multiplication, ordered by difficulty factor

(from 0. 6à:15creau, 1969, Fig. 4).

Dansereau went on to construct a more refined model, in which he postulated

several kinds of memories with associated transfer times between and within

memories. There was an image store, where operands had to be positioned, as in a template, in order to be added or multiplied. There was a short term memory that

held a small number of digits, e.g., the definition of the problem or intermediate

results. Finally, there was a long term memory in which information could be fixated for an indefinite period of time. The transfer times are shown in Figure 6. They were obtained from independent experiments, either already in the literature or done by Dansereau. Thus, these times are not parameters to be estimated from the primary data on performance.

Figure 7 shows the results of this model.

The system is complex enough to

require simulation. The times taken by the simulation are shown as open circles and the actual times by the solid circles. Both are plotted against the difficulty factor used in the prior figure. (Thus there are many dots for a given difficulty

a factor, since there are many different multiplication problems with the same factor.)

It can be seen clearly that the simulation has provided a next order of improvement, fitting the "staircase" effect of the actual data.' This fit is not due to an excess of parameters, since the only parameter used to fit the data was a scale change. All others, as remarked above, were estimated independently from other

data. .

Although we have no space to discuss it, the model shows that very little

time is spent in the act of multiplying or adding. Rather, significant amounts of time are spent in memorizing intermediate results (which we expected) and in positioning operands (which we did not expect).

This work shows clearly the shift from models of memory to models of the

fermediate processor. Memory, of course, remains central to the system, but there is much more as well. Furthermore, we have moved to where an explicit theory must be built of the situation (the simulation), even though the task is still not one

that artificial intelligence finds of much interest per se.

[blocks in formation]

FIGURE 6: Memory transfer rates in mental multiplication

model (after D. Dansereau, 1969, Table 3).

[merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][ocr errors][merged small][merged small][merged small][merged small][ocr errors][merged small][merged small][merged small][merged small][merged small][merged small][merged small][ocr errors][merged small][ocr errors][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small]

FIGURE 7: Average reaction time for mental multiplication for subject TM (*) and Simulation (0)

(from D. Dansereau, 1969, Fig. 19).

6. POINT THREE: ON BEING SERIOUS

I have tried to illustrate with examples from one area, immediate memory, that

theories of man as an information processor are being used in serious and detailed

ways. I would now like to turn this conclusion around. There have always been two feelings held by workers in artificial intelligence about themselves: (1) they were proceeding independently of any concern with human behavior (i.e., not simulating); alternatively (2), they were in fact being relevant to how man thinks. Both these views are, in my mind, legitimate including their conjunction, which has been my personal position in some of our work (e.g., GPS).

I wish to address myself to those of the second (simulating) persuasion. By now, anyone who is serious about the psychological relevance of his work in artificial intelligence had better be prepared to deal with detailed data of humans in specific situations, experimental or otherwise. As we discussed in connection with Figure 1, there are many ways in which a work in artificial intelligence could be considered relevant to the study of human behavior. All these ways remain

legitimate.

But the gradual success of the detailed, use of information processing

theories means that none of the less demanding ways carry much punch (though there

will always be exceptions, naturally).

This same point was reached some years ago with respect to neural modeling and physiology. No neural modeling is of much interest anymore, unless it faces the

detail of real physiological data.

The novelty and difficulty of the tasks under

taken by heuristic programming has tended to push the corresponding day of reckoning

off by a few years. The development of symbolic systems that would behave in any way intelligently produced sufficiency analyses that were in fact relevant to the

psychology of thinking.

But the automatic relevance of such efforts seems to me

about past.

Let me illustrate this point briefly. In the last few years Ross Quillian has developed a model of semantic memory (Quillian, 1965, 1969). Many of you are

undoubtedly aware of it; Bob Simmons discussed it to some extent in his paper at

this conference. The essential features are (1) each concept is a node in a

« PreviousContinue »