How the Stroop Effect Arises from Optimal Response Times in Laterally Connected Self-Organizing Maps

TL;DR

This paper shows that the Stroop effect can arise naturally from the internal dynamics of cortically-inspired, self-organizing lexical and semantic representations. By extending the BiLex SOM with lateral connections, the model produces time-dependent competition between semantic and lexical pathways, with interference regulated by a context-dependent routing parameter $r_\mathrm{lex}$ and an entropy threshold $E_t<1$. The best routing ($r_\mathrm{lex}=0.45$, $r_\mathrm{sem}=0.05$) yields an overall accuracy of $84.2\%$, with a pronounced Stroop effect: congruent trials are fastest and most accurate, incongruent trials are slower and more error-prone, and no-input trials show intermediate performance; reaction-time patterns align with human data. The work highlights how attentional constraints and stimulus-driven processing contribute to cognitive control phenomena and provides a foundation for future extensions to aging, bilingualism, and adaptive routing in cognitive systems, including neural-inspired mechanisms akin to basal ganglia control. $E_t$ and $r_\mathrm{lex}$ are central to the balancing of speed and accuracy, illustrating how optimized performance can incur Stroop-like interference as a natural byproduct of efficient processing.

Abstract

The Stroop effect refers to cognitive interference in a color-naming task: When the color and the word do not match, the response is slower and more likely to be incorrect. The Stroop task is used to assess cognitive flexibility, selective attention, and executive function. This paper implements the Stroop task with self-organizing maps (SOMs): Target color and the competing word are inputs for the semantic and lexical maps, associative connections bring color information to the lexical map, and lateral connections combine their effects over time. The model achieved an overall accuracy of 84.2%, with significantly fewer errors and faster responses in congruent compared to no-input and incongruent conditions. The model's effect is a side effect of optimizing response times, and can thus be seen as a cost associated with overall efficient performance. The model can further serve studying neurologically-inspired cognitive control and related phenomena.

Figures (7)

• Figure 1: The BiLex model of the Stroop effect. The semantic color input and concurrent lexical input are combined in the lexical map, and a lexical output is generated naming the color (maroon in this case). When the inputs are congruent, the response is fast and reliable; when they are incongruent, slower and error-prone. The errors can be seen as a cost of optimization of overall response times.
• Figure 2: The organization and activation in the BiLex model of the Stroop Effect. The model implements SOMs with lateral connections, with map units representing 16 colors and words. The semantic map (top) was trained with RGB values of each color, and the lexical map (bottom) with Spanish phonetic representations of words for each color. Initial map activation for the color maroon/granate is depicted in the right column. Similar colors are nearby in the semantic map, and similar words in the lexical map. As a result, the map activations are smooth and continuous.
• Figure 3: Finding the best values for the $r_\mathrm{lex}$ routing parameter. The normalized $e_i$ (red), $t_i$ (purple), and $t_c$ (green) components of the $q$ metric are depicted for each possible $r_\mathrm{lex}$ value. The optimal routing parameter was found to be an $r_\mathrm{lex}$ of 0.45, marked by the dashed black line, balancing color naming speed with minimized naming errors.
• Figure 4: The mean and standard error (SE) of RTs across the three conditions of the Stroop effect: congruent (red), no-input (blue), and incongruent (green). RTs improve significantly for congruent vs. no-input vs. incongruent conditions, indicating that the model exhibits a strong Stroop effect.
• Figure 5: The mean and SE of RTs in the model compared with human performance. Left: n = 98, ages: 37 - 55 years forte2024stroopdistribution; right: n = 77, ages: 17 - 45 years Wright2017. The vertical scales were adjusted to line up the averages; the SEs are mostly similar and the differences between the conditions are significant, suggesting that the model replicates the human data.