Experience and Pseudo-Experience: Exemplar Effects Without Feedback

Many real world situations do not offer unambiguous outcome feedback on how to categorize objects. Models in the categorization literature have mostly been formulated for tasks with trial-by-trial outcome feedback. We examined if there was evidence for exemplar memory also when no external feedback is provided and the criterion is derivative of more abstract knowledge. In a “teacher-student” task, a teacher learns how to judge the toxicity of bugs from external outcome feedback and conveys this knowledge to a student that receives no outcome feedback. The results showed that the students exhibit exemplar effects even if the instructions from the teachers were in the form of rules. Introduction Consider listening to your very first speech by a politician. Your previous knowledge is likely to influence your attitude towards her or him. Perhaps, already from childhood your father has imprinted in you that politicians are guided by strictly egoist motives and your general conceptions thus include a belief that no politician advocates a proposal that does not lay in his or her personal interest. You hear a short speech that is neutral in content. Later you listen to another politician. How is your opinion of this second politician influenced by the first encounter? You did not receive much useful feedback from the first encounter, as you only listened to a short neutral speech. However, in your memory the first politician is stored as a person only interested in pursuing his own interest. This exemplar memory only in part derives from direct experience with a politician; in part it is derivative of more general beliefs held prior to the encounter. However, by now this belief is supported also by “concrete experience” of politicians. This was an example of a real world situation that is different from most categorization experiments where classification models are tested in tasks with simple perceptual stimuli and trial-by-trial outcome feedback. Everyday situations often do not offer direct and unambiguous feedback and exemplar memory is thus likely to derive in part also from other sources of knowledge besides concrete experience with the objects. There has been increasing interest in multiple representation levels (e.g., Ashby, Alfonso-Reese, Turken, & Waldron, 1998) and there is evidence that people can adaptively shift between different representation levels in response to the experimental demands (Jones, Juslin, Olsson, & Winman, 2000). With experience, knowledge first represented abstractly may be projected onto concrete exemplars so that in the end the beliefs are supported also by an extensive exemplar memory: a phenomenon that might be called pseudo-experience. Even if some extensions of exemplar models allows for storage of exemplars as they are interpreted and not solely in terms of their physical properties (e.g., the model presented by Smith & Zarate, 1992), the argument supporting this claim is based on general observations and not linked to predictions from different models. For example, one such observation is that a reencounter of a stimulus facilitate the same reactions and processes (see the review in Smith & Zarate, 1992). In this paper, we examine the possibility of extending the scope of exemplar models (Medin & Schaffer, 1978; Nosofsky, 1986) to situations where people do not receive outcome feedback, but form beliefs about the criterion from abstract knowledge of rules. In these circumstances, one possibility is that people completely abandon exemplar processes as a basis for their judgments and directly use abstract knowledge in the form of rules or prototypes. Another possibility is that people generate the criteria from abstract knowledge and store them together with the experienced exemplars; later to rely on these stored exemplars to make their judgments. We explore these possibilities in a “teacher-student” task where a teacher learns how to judge the toxicity of bugs from outcome feedback and the student has to rely on feedforward summary information provided by the teacher. The question is if there is evidence for exemplar processing in the students judgments even if they do not receive feedback or instructions about exemplars from the teachers. Measuring Exemplar Effects To develop an exemplar effect index we need to consider a category structure that allows us to differentiate between predictions by the exemplar model and other plausible models, in this case a cue-abstraction model that linearly integrates cues. The results previously obtained with this task revealed large individual differences and a shift from exemplar memory to more mental cue-integration processes when the criterion is changed from classification to a continuous judgment task (Juslin, Olsson, & Olsson, 2002). The task requires participants to use four binary cues to infer a continuous criterion. (Juslin et al., 2002). The judgments involve the toxicity of subspecies of a fictitious Death Bug. The different subspecies vary in concentration of poison from 50 ppm (harmless) to 60 ppm (lethal). The toxicity can be inferred from four visual cues of the subspecies (e.g., the length of their legs, color of their back). The binary cues C1, C2, C3, and C4 take on values 1 or 0. The toxicity c of a subspecies is a linear, additive function of the cue values: 4 3 2 1 1 2 3 4 50 C C C C c ⋅ + ⋅ + ⋅ + ⋅ + = . (1) C1 is the most important cue with coefficient 4 (i.e., a relative weight .4), C2 is the second to most important with coefficient 3, and so forth. A subspecies with feature vector (0, 0, 0, 0) thus has poison concentration 50 ppm; a subspecies with feature vector (1, 1, 1, 1) has 60 ppm. The continuous criteria for all 16 subspecies (i.e., possible cue configurations) are summarized in Table 1. Table 1 Structure of the Task Exemplar C1 C2 C3 C4 Criterion Exemplartype 1 1 1 1 1 60 E 2 1 1 1 0 59 T 3 1 1 0 1 58 T 4 1 1 0 0 57 O 5 1 0 1 1 57 N 6 1 0 1 0 56 N 7 1 0 0 1 55 N 8 1 0 0 0 54 T 9 0 1 1 1 56 O 10 0 1 1 0 55 O 11 0 1 0 1 54 T 12 0 1 0 0 53 T 13 0 0 1 1 53 T 14 0 0 1 0 52 T 15 0 0 0 1 51 T 16 0 0 0 0 50 E Note: C = Cue; E = Extrapolation; T = training exemplar; O = Old comparison exemplar in training, N = New comparison exemplar presented at test. In training, the participants encounter 11 subspecies and make continuous judgments about the toxicity of each subspecies (“The amount of poison is 57%”). As indicated in the two right-most columns of Table 1, five subspecies are omitted in training. In a test phase, the participants make the same judgments as in the training phase, but for all the 16 subspecies and without feedback. The task allows perfect performance in training both by exemplar memory and induction of the task structure (i.e., by inducing the cue weights in Eq. 1). The exemplar model implies that participants make judgments by retrieving similar exemplars (subspecies) from long-term memory. The context model of classification (Medin & Schaffer, 1978) suggests that in a task that only requires participants to judge if a bug is dangerous or not, the probability ) 1 ˆ ( = b pE of categorization as dangerous (1) equals the ratio between the summed similarity of the judgment probe to the dangerous exemplars and the summed similarity to all exemplars: