Across three different conditions, participants learned to classify images of rocks into the high-level categories of igneous, metamorphic and sedimentary. In all conditions, the instructions to the participants emphasized that their primary task was to learn these high-level category assignments. In some of the conditions, the participants also learned to classify the rocks into their subtype categories. The complete set of rocks comprised 30 subtypes, 10 subtypes from each of the three high-level categories. The subtypes are listed in Table 1. The subtypes are highly representative of those that are commonly taught in introductory college-level geoscience classes, and are among the major ones listed and described in introductory textbooks (e.g., Marshak, 2015; Tarbuck & Lutgens, 2017). Because it was unrealistic to expect a participant to learn all 30 subtypes in a single 1-h session, each individual participant was randomly assigned 15 of the 30 subtypes to learn (five from each of the three high-level categories).Footnote 2
In condition 1, participants were trained on only the high-level names of the rocks. An example screenshot of the question prompt on a typical trial is presented in Fig. 2. As illustrated, on each trial, an individual rock would be presented, and the participant would attempt to classify it into one of the three high-level categories. Feedback was provided only with respect to the high-level category to which the rock belonged. In the test phase, participants continued to classify items at only this high-level of categorization.
In condition 2, participants learned simultaneously to classify rocks into both their high-level and subtype-level categories. The condition used a two-stage response procedure. An example screenshot of the first stage of an individual trial is presented in Fig. 3. As illustrated, an individual rock would be presented in the center of the screen. Underneath the rock, the high-level responses igneous, metamorphic and sedimentary were shown in three columns, and beneath each high-level name were shown the subtypes for that high-level category. In the first stage, participants were prompted to enter the high-level response for the rock. Once the high-level response was selected, the second response stage began. As shown in the example screenshot in Fig. 4, the participant was prompted to select the subtype name from among the possibilities for the selected high level. For instance, if a participant had responded that the rock was metamorphic, he or she would then be prompted to select the rock’s subtype name within the metamorphic column. This same two-stage procedure for collecting responses continued to be used in the testing phase of condition 2.
Our central idea in implementing the two-stage response procedure of condition 2 was that it might combine in synergistic fashion various elements of previously tested procedures that have advantageous components. First, because participants are required initially to classify at the high level, they may be motivated to search for features that are diagnostic at that high level. Second, the requirement that participants also learn the subtype-level categories may foster the learning of aspects of the category structure that are disorganized and dispersed (e.g., in which subtypes from contrasting high-level categories are highly similar to one another). Third, the requirement that participants make two separate responses on each individual trial — first the high-level classification response and then the subtype-level one — might be effective in allowing participants to develop learned associations between the high-level and subtype-level names of the rocks.
Nevertheless, in terms of assessing the participants’ acquired knowledge, this experimental condition has the same limitation as did the simultaneous paired-name condition that had been tested in Nosofsky et al.’s (2017) experiment. In particular, because the high-level and subtype-level names were simultaneously present, a participant could in principle focus on only the subtype names during both training and test. On each trial, if the participant decided that a rock was, for instance, ‘granite’, then he or she could enter the corresponding high-level category response (‘igneous’) by making reference to the column in which granite appeared. Thus, an alternative condition was required to evaluate the extent to which the training procedure is effective in allowing participants to directly classify the rocks at the high-level of categorization.
We addressed this requirement by also conducting condition 3. With one exception described below, the training phase for condition 3 was identical to that in condition 2. The key difference across the conditions arose at time of test. Whereas in condition 2 we continued to present the subtypes along with the high-level names at the time of the test (in the column format illustrated in Figs. 3 and 4), in condition 3 the subtypes were no longer presented. Instead, just as in condition 1, the question prompt now made reference to only the high-level categories (as illustrated in Fig. 2). Thus, condition 3 provided a pure test of the participants’ ability to classify the rocks into the high-level categories, without the benefit of an external cue that linked the subtype names to the high-level names.
The second difference between conditions 2 and 3 arose during the training phase. On 80% of the trials, the same two-stage response procedure was used in condition 3 as in condition 2. However, on 20% of the trials, the question prompt was the same as in condition 1; that is, participants were required to classify the rock into one of the high-level categories without the benefit of the external cue showing which subtype names were linked with which high-level names (as in Fig. 2). In addition, on these trials, a reminder message was provided at the bottom of the computer screen stating: “Remember, your primary job is to learn the high-division names.” We included these high-level-only trials to remind participants that their primary task was to learn the high-level name for each rock and to discourage participants from developing a strategy of relying solely on learning the subtype-level names.
A schematic summary of the training and testing procedures across the three conditions is provided in Fig. 5.
Method
Participants
There were 95 undergraduate students from Indiana University Bloomington who participated as part of a requirement for their introductory psychology courses. The participants all had normal or corrected-to-normal vision and all reported having normal color vision. All reported that they had little or no previous experience in rock classification. Each participant’s condition was randomly assigned, with 32 participants in condition 1, 32 in condition 2, and 31 in condition 3. These sample sizes were as large or larger than in the individual conditions of the closely related studies that most directly motivated the present research and that found significant differences in the outcomes of the broad- versus specific-level training (Miyatsu et al., in press, experiments 1 and 2; Noh et al., 2014; Nosofsky et al., 2017). (As it turned out, the correlation on the repeated old–new item performance measure in our study was r = .74; this yielded power = .628 to detect a medium-size main effect of training procedure on test-phase performance, and power = .966 to detect a large-size effect.)
Stimuli and apparatus
The stimuli consisted of 360 pictures of rocks from the three broad divisions of igneous, metamorphic and sedimentary rocks. Each broad division comprised 10 subtype categories, listed in Table 1. There were 12 samples of each subtype. The rock picture samples were taken from a variety of online sources (for a fuller description of these stimulus materials, see Nosofsky, Sanders, Meagher, & Douglas, 2018). The experiment was programmed in MATLAB using Psychophysics toolbox (Brainard, 1997) on a personal computer running Microsoft Windows.
Procedure
For each individual participant, 5 of the 10 subtypes from each of the three broad divisions of igneous, metamorphic and sedimentary rocks were randomly selected. The participant learned the items from only these randomly selected subtypes. We used this procedure of sampling a subset of the categories because pilot work suggested that overall learning performance would be poor if participants were required to try to learn all 30 categories in a single 1-h session.
In all conditions, the procedure consisted of a training phase and a test phase. The training phase consisted of three training blocks with 90 trials in each block, whereas the test phase consisted of one block with 120 trials. The test phase included presentations of old items from the training phase as well as novel transfer items from the studied categories. Across all conditions, for each individual participant, the members of each rock subtype were randomly assigned as either training or novel transfer stimuli. For each subtype, there were six randomly chosen training examples and four randomly chosen transfer examples.
Across all conditions in the training phase, on each trial, a picture of a training rock was displayed in the center of the screen and the participant attempted to classify it. Each of the individual training items was presented once per block, with the order of presentation of the 90 training items randomized. An analogous procedure was used across all conditions in the test phase. The tested stimuli consisted of four of six randomly selected training examples from each of the 15 subtype categories, and of the four novel transfer items from each of the 15 subtype categories, for a total of 120 test items. The order of presentation of the 120 items was randomized.
The nature of the training and test procedures in each of the conditions has already been described in our introduction to this experiment; here we provide only some additional methodological details. First, in cases in which participants were required to classify the rocks into their high-level categories, they did so by pressing the “i” key for igneous, “m” for metamorphic and “s” for sedimentary. In cases in which participants were required to indicate the subtype category of the rock, they did so by pressing a number on the keyboard that preceded the subtype name on the computer screen (as illustrated in Fig. 4). During the training phase, the computer displayed corrective feedback at the end of each trial (with the rock picture remaining on screen), stating that the participant was either “correct” or “incorrect” followed by the correct response. In condition 1 (and in the 20% of trials in condition 3 that required only a high-level response), the corrective feedback was with respect to only the high-level category of the rock (for example, “Correct! Igneous” or “Incorrect: Sedimentary”). In condition 2 (and in the 80% of trials in condition 3 that used the two-stage response procedure), the corrective feedback was provided after both responses were made. The computer provided simultaneous feedback at both levels, such as “Correct! Igneous-gabbro” or “Incorrect: Metamorphic-marble”. In all conditions, the feedback remained on the screen for 1 s following correct responses and for 2 s following incorrect responses (with the picture of the rock remaining on the screen). The feedback was followed by a 0.5-s inter-trial interval consisting of a blank screen. At the end of each training block, the computer reported to the participants their overall percentage of correct responses. The methodological details of the test phase were the same as already described for the training phase, except no corrective feedback was provided during the test-phase trials. Instead, the computer simply displayed a message of “okay” to indicate to the participants that their response had been recorded. At the end of the testing phase, the participants were thanked and were provided with a debriefing of the purpose of the experiment. The experimental session lasted roughly 50 min.
Results
Training
To analyze the results from the training phase, we divided the complete sequence of 270 trials into 18 consecutive 15-trial sub-blocks and then measured the mean proportion of correct high-level responses in each sub-block. The results for each of the three conditions are shown in Fig. 6. These data were submitted to a 3 × 18 mixed model analysis of variance (ANOVA), with condition (1, 2, 3) as the between-subjects variable and sub-block (1–18) as the within-subjects variable. As can be seen, overall performance improved dramatically in all three conditions as a function of training, F (17, 1564) = 49.014, MSE = .016, p < .0001; the degree of improvement in performance did not significantly vary across conditions, F (34, 1564) = .445, p = .998 for the interaction. Thus, not surprisingly, our naive participants did not enter the experiment with significant amounts of prior knowledge of the high-level rock category assignments, but rather learned these assignments during the course of the training phase.
Importantly for present purposes, there was no significant difference in high-level naming performance as a function of condition, F(2, 92) = .934, MSE = .148, p = .397. Thus, requiring participants to make a subtype-level classification following the high-level classification during training (conditions 2 and 3) did not negatively impact learning of the high-level classification of the training items relative to focusing participants only on high-level classification (condition 1); indeed, if anything, learning of high-level classifications was slightly better when subtype classification also had to be learned (see Fig. 6).
Test
The results from the test phase are shown in Figs. 7 and 8. The figures show the mean proportion with which participants classified members of each main high-level division of rocks (igneous, metamorphic, sedimentary) into each of the high-level divisions. Within each figure, the top panels show the results for the old training items whereas the bottom panels show the results for the new transfer items. For ease of comparison, Fig. 7 shows performance for condition 1 (left panels) versus condition 2 (right panels), whereas Fig. 8 shows performance for condition 1 (left panels) versus condition 3 (right panels).
Inspection of the results reveals that participants classified the rocks into their correct high-level categories well above chance levels (.33) in each of the conditions. In addition, as would be expected, correct classification performance for the old training items presented during the test phase was higher than for the novel transfer items. The pattern of results was also fairly similar across each of the three main divisions of rocks (igneous, metamorphic and sedimentary).
The key question concerns the comparisons of performance across the different training and testing conditions. Participants in condition 2 achieved nominally higher correct proportions than did participants in condition 1 on both the old training items (condition 1, M1 = .70, SD1 = .14; condition 2, M2 = .73, SD2 = .12) and the new transfer items (condition 1, M1 = .61, SD1 = .12; condition 2, M2 = .64, SD2 = .10). Participants in condition 3 performed virtually the same as did participants in condition 1 on both the old training items (M1 = .70; M3 = .70, SD3 = .15) and the new transfer items (M1 = .61; M3 = .62, SD3 = .13). We analyzed these data with a 3 × 2 mixed-model ANOVA, with condition (1–3) as the between-subjects factor and item type (old versus new) as the within-subjects factor. Old items were classified significantly more accurately than new items, F(1, 90) = 93.91, MSE = .004, p < .001. However, there was no significant effect of condition, F(2, 90) = 0.46, MSE = .030, p = .63, and no significant interaction between condition and item type, F(2, 90) = 0.17, MSE = .004, p = .85.
Discussion
In their previous study, Miyatsu et al. (in press) found that direct high-level-only training led to significantly better high-level classification performance at the time of the test than did training in which participants were required to also learn the subtype-level names of the rocks (see also Noh et al., 2014). If anything, our results go slightly in the opposite direction, with performance in the simultaneous paired-name conditions being at least as good as performance in the high-level-only condition. Importantly, this pattern held even in the paired-name training condition in which participants were no longer provided with the subtype names of the rocks at the time of transfer. However, before discussing the most likely reasons for these contrasting results, we first report a second study to confirm the reliability of our findings. Our tentative conclusion from the present experiment is that the present form of paired high-level/subtype-level training leads to no disadvantage in high-level naming performance compared to the case in which participants are trained at only the high level. Because the conclusion rests on a finding of no difference (i.e., a null result), a possible concern is that forms of experimental noise could be hiding a high-level-only advantage. One such potential form of experimental noise is that each participant in experiment 1A was exposed to randomly selected subsets of rock subtypes from each of the high-level categories. If, by happenstance, participants in the paired-name condition were exposed to an easier set of subtypes, then the results could be hiding what is a true high-level-only advantage. We conducted experiment 1B to address this possibility.