Speeded Classification
Finally, after the subjects had completed the transfer task, they were asked to perform the classification task again with the original nine faces, but respond as rapidly as they could. In Table IV, we show the average reaction times of the subjects in responding to each face for each set of instructions, and compare these with EPAM's reaction times, without modifying any of EPAM's time parameters from their usual values.
In the three conditions, the subjects took, on average, 31, 28, and 24% longer than the EPAM simulation. Hence, the times predicted with parameters obtained from earlier studies of rote verbal learning provided a reasonable fit to the data. There is a high rank-order correlation between the times, averaged over subjects, taken to respond to the different faces and the numbers of errors they had made while learning the faces. Custom Admission essay writing services and educated admission essay writers provide the best assistance to students. In the three experimental conditions, there are rank-order correlations between EPAM's times on individual faces and the subjects' times of .71, .50, and .13, respectively. Thus, the speeded-classification task shows much the same pattern of findings as the two previous tasks.
In this blog we have described how EPAM, a program originally constructed to predict the behavior of human subjects in verbal learning experiments, can be used to predict behavior in categorization experiments, with-out the need to modify substantially the basic learning and performance mechanisms of the system or the time parameters that predict rate of learning and speed of response. To illustrate how EPAM accomplished this, we took as an example a task that had been studied by Medin and Smith under three different conditions that corresponded to three different sets of task instructions.
This task is of special interest because it requires subjects to use a different strategy for each of the experimental conditions. Hence, the data of the subjects' performance reflect not only their own learning and response capabilities, but also differences of difficulty in categorizing the individual stimuli (characteristics of the task domain) and differences in the strategies they adopt. Although the importance for task difficulty of the task domain, the subjects' representation of the task domain (the problem space), and the strategies employed by subjects has been known for a long time, there are still relatively few published experiments in which these variables have been manipulated, or in which the subjects' behavior on these dimensions have been recorded and reported.
In the three conditions, the subjects took, on average, 31, 28, and 24% longer than the EPAM simulation. Hence, the times predicted with parameters obtained from earlier studies of rote verbal learning provided a reasonable fit to the data. There is a high rank-order correlation between the times, averaged over subjects, taken to respond to the different faces and the numbers of errors they had made while learning the faces. Custom Admission essay writing services and educated admission essay writers provide the best assistance to students. In the three experimental conditions, there are rank-order correlations between EPAM's times on individual faces and the subjects' times of .71, .50, and .13, respectively. Thus, the speeded-classification task shows much the same pattern of findings as the two previous tasks.
In this blog we have described how EPAM, a program originally constructed to predict the behavior of human subjects in verbal learning experiments, can be used to predict behavior in categorization experiments, with-out the need to modify substantially the basic learning and performance mechanisms of the system or the time parameters that predict rate of learning and speed of response. To illustrate how EPAM accomplished this, we took as an example a task that had been studied by Medin and Smith under three different conditions that corresponded to three different sets of task instructions.
This task is of special interest because it requires subjects to use a different strategy for each of the experimental conditions. Hence, the data of the subjects' performance reflect not only their own learning and response capabilities, but also differences of difficulty in categorizing the individual stimuli (characteristics of the task domain) and differences in the strategies they adopt. Although the importance for task difficulty of the task domain, the subjects' representation of the task domain (the problem space), and the strategies employed by subjects has been known for a long time, there are still relatively few published experiments in which these variables have been manipulated, or in which the subjects' behavior on these dimensions have been recorded and reported.
Transfer Task
After completing their initial learning of the classification of the nine faces, subjects were given a transfer task, in which they were asked to categorize the same faces again, intermingled with examples of seven new faces. The results of the transfer experiment are shown in Table III.
Again, there is a close relation between the subjects' data and the EPAM simulations on the transfer test, the relation being somewhat closer for the old than for the new faces. EPAM tends to move closer to chance (50%) on the new faces, which is consistent with the fact that it sometimes guessedthe categories correctly in the learning experiment without extending the differentiation net to classify them unambiguously. If you need custom written essays, purchase college essay writing services at American site! Nevertheless, the correlation over all conditions combined between errors made by EPAM and errors made by subjects for all the different faces was .82, but .88 for the faces seen previously and .73 for the seven new faces.
Medin and Smith used regression models (a "context" or multiplicative model, and an "independent cue" or additive model) to fit the data from the transfer experiment, obtaining average absolute deviations about a third as large as EPAM's in the first case, and about half as large in the second. However, each of these models had 4 free parameters that were used in fitting the data, and were estimated separately for the three experimental conditions -- a total of 12 parameters. Hence, it is hard to conclude that the regression models did a better job than EPAM of fitting the facts. Medin and Smith remark that the results were very sensitive to the exact values of the parameters, which suggests that the parameters were doing much of the work. This kind of flexibility was not available to EPAM, for the same interpretations of the instructions were used to model both the learning and transfer experiments.
Again, there is a close relation between the subjects' data and the EPAM simulations on the transfer test, the relation being somewhat closer for the old than for the new faces. EPAM tends to move closer to chance (50%) on the new faces, which is consistent with the fact that it sometimes guessedthe categories correctly in the learning experiment without extending the differentiation net to classify them unambiguously. If you need custom written essays, purchase college essay writing services at American site! Nevertheless, the correlation over all conditions combined between errors made by EPAM and errors made by subjects for all the different faces was .82, but .88 for the faces seen previously and .73 for the seven new faces.
Medin and Smith used regression models (a "context" or multiplicative model, and an "independent cue" or additive model) to fit the data from the transfer experiment, obtaining average absolute deviations about a third as large as EPAM's in the first case, and about half as large in the second. However, each of these models had 4 free parameters that were used in fitting the data, and were estimated separately for the three experimental conditions -- a total of 12 parameters. Hence, it is hard to conclude that the regression models did a better job than EPAM of fitting the facts. Medin and Smith remark that the results were very sensitive to the exact values of the parameters, which suggests that the parameters were doing much of the work. This kind of flexibility was not available to EPAM, for the same interpretations of the instructions were used to model both the learning and transfer experiments.
Comparison with Human Data
A single free parameter called the "study parameter" was adjusted so that approximately the same proportion of EPAM-simulated subjects as real subjects would attain the same overall result in the learning condition. Medin and Smith reported that 14 of 32 people met the criterion of a perfect trial in both the standard and rules-plus-exceptions conditions, while only 8 of 32 people met the criterion in the prototype condition. With the study parameter set at 18, so that EPAM studied only 18% of the time when studying was possible, 151 of 320 simulated subjects met the criterion in the standard condition, 129 of 320 met the criterion in the rules-plusexceptions condition, and 63 of 320 met the criterion in the prototype condition, closely matching the ratios for the human subjects. Cheap college essay for college students delivered by reliable college essay writers online. Both EPAM and people found it easier to meet the criterion in the standard and rulesplus-exceptions condition than in the prototype condition.
There was no special adjustment of parameters for this simulation, and simple and straightforward interpretations of the instructions were used for EPAM. For all three experimental conditions, the Pearsonian correlation between human subjects and EPAM of the numbers of errors for the various faces is very high: .93, .93, and .77 for the three conditions.
In EPAM as in the human experiments, Faces 13, 2, and 12 (except 12 in the Rules and Exceptions condition) produced by far the largest number of errors. There was little difference in rank order among the several conditions for either the human subjects or EPAM. In this important respect, the instructions had little effect on the outcomes, affecting only the overall level of difficulty of the task as a whole. The relative reduction in difficulty of Face 12 in the Rules and Exceptions condition was reflected in the performance of both subjects and EPAM.
Face 7 produced fewer errors than Face 4 in all conditions, which, as Medin and Smith point out, is consistent with context models but not with independent-cue models, where the net effect of cues is additive. As EPAM is, in many respects, highly nonlinear and nonadditive in its operation, hence a "context" model in the sense of Medin and Smith, we would predict this result. On average, both the subjects and EPAM found the prototype condition hardest, the standard condition next hardest, and the rules-andexceptions condition easiest.
The range of errors from the easiest to the most difficult faces was smaller for EPAM than for the human subjects in the standard and prototype conditions, but not in the rules-plus-exception conditions.
There was no special adjustment of parameters for this simulation, and simple and straightforward interpretations of the instructions were used for EPAM. For all three experimental conditions, the Pearsonian correlation between human subjects and EPAM of the numbers of errors for the various faces is very high: .93, .93, and .77 for the three conditions.
In EPAM as in the human experiments, Faces 13, 2, and 12 (except 12 in the Rules and Exceptions condition) produced by far the largest number of errors. There was little difference in rank order among the several conditions for either the human subjects or EPAM. In this important respect, the instructions had little effect on the outcomes, affecting only the overall level of difficulty of the task as a whole. The relative reduction in difficulty of Face 12 in the Rules and Exceptions condition was reflected in the performance of both subjects and EPAM.
Face 7 produced fewer errors than Face 4 in all conditions, which, as Medin and Smith point out, is consistent with context models but not with independent-cue models, where the net effect of cues is additive. As EPAM is, in many respects, highly nonlinear and nonadditive in its operation, hence a "context" model in the sense of Medin and Smith, we would predict this result. On average, both the subjects and EPAM found the prototype condition hardest, the standard condition next hardest, and the rules-andexceptions condition easiest.
The range of errors from the easiest to the most difficult faces was smaller for EPAM than for the human subjects in the standard and prototype conditions, but not in the rules-plus-exception conditions.
Prototype Instructions
Medin and Smith's instructions for their "prototype" subjects were to memorize what A faces look like and what B faces look like. They were told that they later would have to answer questions about the characteristics of each type of face. EPAM memorizes types of faces by making a separate net for each type. Medin and Smith's instructions were: "we want you to use these general impressions to help you classify these faces." EPAM's find-category routine for the prototype condition does this by sorting each stimulus in both nets. If a face is found to be a member of one category but not the other, EPAM responds with the former category. Customized Admission writing services and trained admission essay writers offer expert help to students. If it is found to be a member of both or is not found to be a member of either, the subject module guesses the category. If it guesses wrong, it elaborates the net for the correct category in which to include this stimulus.EPAM's study-category routine for the prototype condition follows almost the identical strategy as the study-category routine for the standard condition, except that there is the additional step: the study-category routine must determine which net to use for studying. Specifically:
1. If the find-category routine responds correctly, do nothing.
2. Pick a random number from 0 to 99, if it is over 17, do nothing.
3. If a previous study-category routine is busy transferring information to long-term memory, do nothing.
4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensory store) share the same response and at least three features in common, form a generalization consisting of the features common to both stimuli. If the stimulus was not identified as a member of the correct net, then memorize the generalized stimulus completely in the correct net and associate it with "Yes" in that net. On the other hand, if the stimulus was identified as a member of the correct net, then it must have been misidentified as a member in the other net, so memorize the generalized stimulus completely in the other net and associate it with "No" in that net.
5. If the stimulus was not identified as a member of the correct net then associate it with "Yes" in that net. If the stimulus was identified as a member of the correct net, then memorize the stimulus completely in the other net and associate it with "No" in that net.
1. If the find-category routine responds correctly, do nothing.
2. Pick a random number from 0 to 99, if it is over 17, do nothing.
3. If a previous study-category routine is busy transferring information to long-term memory, do nothing.
4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensory store) share the same response and at least three features in common, form a generalization consisting of the features common to both stimuli. If the stimulus was not identified as a member of the correct net, then memorize the generalized stimulus completely in the correct net and associate it with "Yes" in that net. On the other hand, if the stimulus was identified as a member of the correct net, then it must have been misidentified as a member in the other net, so memorize the generalized stimulus completely in the other net and associate it with "No" in that net.
5. If the stimulus was not identified as a member of the correct net then associate it with "Yes" in that net. If the stimulus was identified as a member of the correct net, then memorize the stimulus completely in the other net and associate it with "No" in that net.
EPAM's Study-Category Routine for the Rules-Plus-Exceptions Condition
EPAM's study-category routine for the rules-plus-exceptions condition is also a two-stage algorithm. In the first stage, if the subject has not yet learned the rule, it forms a hypothesis, and studies the rule following a strategy very much like that outlined in the instructions. Pay for Writing service by professional writing experts and you'll receive properly written essay within the shortest deadline! Specifically, it accepts a nose-length rule, such as "short nose predicts Category A," when its excess of correct over incorrect predictions exceeds 3, and accepts the opposite nose-length rule (i.e., "short nose predicts Category B") when the excess of correct over incorrect predictions dips below zero.The second stage is very much like the find-category routine except that the system is determining whether or not the face is a member of the exceptions net. Specifically:
1. If the find-category routine responds correctly, do nothing.
2. Pick a random number from 0 to 99, if it is over 17, do nothing.
3. If a previous study-category routine is busy transferring information to long-term memory, do nothing.
4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensory store) share the same response and at least three features in common, form and completely memorize a generalization consisting of the features that are on both stimuli and label that generalization as exception or not, consistently with the present stimulus.
5. Otherwise label the present stimulus as an exception or not, as the case may be.
The two discrimination nets after completion of EPAM's learning stage in a run of the rules-plus-exceptions condition are illustrated in Fig. 5. The net on the left is a rules net with a single test for nose length. The net on the right is an exceptions net with a top test for mouth height. In this case, EPAM went through all 32 learning trials without a perfect trial, and, indeed, using this net, EPAM currently misclassifies both of the exceptions to the rule, Faces 13 and 2.It currently takes EPAM about 960 ms to categorize each face: 10 ms to react to the stimulus, 100 ms to enter each of the two nets, and 250 ms to sort through each of the three tests (the test for NL in the Rule net and the tests for MH and ES in the Exception net.During the speeded-classification phase of the experiment, the system adds additional tests to the net, and these permit it to discriminate the exceptions.
PR: wait... I: wait... L: wait... LD: wait... I: wait...wait... Rank: wait... Traffic: wait... Price: wait... C: wait...
1. If the find-category routine responds correctly, do nothing.
2. Pick a random number from 0 to 99, if it is over 17, do nothing.
3. If a previous study-category routine is busy transferring information to long-term memory, do nothing.
4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensory store) share the same response and at least three features in common, form and completely memorize a generalization consisting of the features that are on both stimuli and label that generalization as exception or not, consistently with the present stimulus.
5. Otherwise label the present stimulus as an exception or not, as the case may be.
The two discrimination nets after completion of EPAM's learning stage in a run of the rules-plus-exceptions condition are illustrated in Fig. 5. The net on the left is a rules net with a single test for nose length. The net on the right is an exceptions net with a top test for mouth height. In this case, EPAM went through all 32 learning trials without a perfect trial, and, indeed, using this net, EPAM currently misclassifies both of the exceptions to the rule, Faces 13 and 2.It currently takes EPAM about 960 ms to categorize each face: 10 ms to react to the stimulus, 100 ms to enter each of the two nets, and 250 ms to sort through each of the three tests (the test for NL in the Rule net and the tests for MH and ES in the Exception net.During the speeded-classification phase of the experiment, the system adds additional tests to the net, and these permit it to discriminate the exceptions.
PR: wait... I: wait... L: wait... LD: wait... I: wait...wait... Rank: wait... Traffic: wait... Price: wait... C: wait...
