Examining the Relationship Between Early Experience, Selective Attention, and the Formation of Learning Traps

Abstract

A simple-rule learning trap occurs when people show suboptimal category learning due to insufficient exploration of the learning environment. By combining experimental methods and computational modeling, the current study investigated the impact of two key factors believed to play essential roles in the development of a simple-rule learning trap: early learning experience and selective attention. Our results showed that, in a learning environment where the true category mapping was determined by conjunctions of two predictive dimensions, the likelihood of falling into a single-dimensional learning trap increased when early learning experience involved a large loss that could be predicted from a single feature dimension. In addition, using a model-based measurement of attention bias, we observed that early experience affected trap formation by narrowing the distribution of attention to exemplar features. These findings provide the first direct empirical evidence of how early learning experience shapes the formation of a simple-rule learning trap, as well as a more granular understanding of the role of selective attention and its interaction with early learning experience in trap formation.

Keywords: Category learning; Exemplar models; Selective attention.

Fig. 1

Stimulus design. Panel (a): Exemplar binary‐valued bee features, including numbers of legs (two vs. six), numbers of wing pairs (single vs. double), and body patterns (striped vs. dotted). Two feature dimensions were randomly selected as relevant features for category membership for each participant, and one was irrelevant. Panel (b): An example choice screenshot where participants make an approach/avoid decision on the present bee by clicking on the respective alternate buttons at the bottom of the screen. The total amount of accumulated points is displayed at the top‐right of the screen.

Fig. 2

Category rules and key manipulations of early learning experience. Panel (a): Schematic optimal two‐dimensional category rule of different bee types (i.e., s00, s01, s10, s11) based on relevant dimensions. 0 denotes a safe feature and 1 denotes a potentially dangerous feature. Combinations of these features determine four bee types. The red dashed line notates the conjunctive category bound. The bee type composed of two potentially dangerous features (i.e., s11) is dangerous and should be avoided. All other bee types are safe to approach. (b): Schematic one‐dimensional category rule on the basis of either relevant dimension 1 (left plot) or 2 (right plot). The dark‐red lines notate the suboptimal one‐dimensional category bound. (c): The manipulated order of early learning sequence (i.e., bee items shown in the first 24 trials of the experiment). In the Learning Trap (LT) Promote condition, participants encounter only one type of the ambiguously safe bee types (i.e., s01) before encountering a dangerous bee. In the LT Prevent condition, participants encounter all safe bee types before encountering a dangerous bee.

Fig. 3

Prevalence of different rule users within each block. Top Row: Proportions of different category‐strategy users across blocks in each early experience condition with full category feedback. Bottom Row: Proportions of different category‐strategy users across blocks in early experience conditions with contingent category feedback. Blocks 1–6 are learning blocks, and Block 7 is the test block. The left y‐axis denotes the proportion for 1D‐rule users, and the right y‐axis denotes the proportion for 2D‐rule users.

Fig. 4

Prevalence of different rule users across blocks in the replication study. Proportions of different category‐strategy users across blocks in early experience conditions (left: LT Promote; right: LT Prevent) with contingent category feedback. Blocks 1–6 are learning blocks, and Block 7 is the test block. The left y‐axis denotes the proportion for 1D‐rule users, and the right y‐axis denotes the proportion for 2D‐rule users.

Fig. 5

Degree of attention bias in each condition. Selective attention to relevant dimensions for each individual was measured by averaging the absolute difference in normalized attention weights between relevant dimensions across trials. α₁ and α₂ are the respective attention weights for relevant dimensions 1 and 2. Since ALCOVE‐RL does not impose any bounds for attention weights, we normalized attention weights by dividing them by the total sum of weights in each trial for each participant for comparison. Error bars denote the standard error of the mean. Brackets denote post‐hoc pairwise comparisons between early sequence conditions with either full or contingent feedback, using Tukey‐adjusted p‐values. ***: p < .001, **: .001 < p ≤ .01, ns: not significant.