Ambiguity in category leaning and stereotype formation

This project will evaluate the ability of Formal Models of Category Learning to explain participants response behaviour across multiple phenomena. We will implement each phenomena as a social categorization problem involving within- and between-group discrimination. We focus on problems, where participants' responses contrast the predictions of a rational approach, like that of Classical Probability Theory. Most of these phenomena also share some other similarities: each has categories occurring at different frequencies during learning and each arises when people generalize what they learned to new, often ambiguous and novel information.
The project currently includes the inverse base-rate effect, illusory correlation, and highlighting. We hope to implement more phenomena as they become replicated or we replicate them ourselves. This research tries to understand the cognitive processes underlying these irrational effects by using Formal Mathematical Models and Open Science techniques.

Compared to natural language explanations, Formal Models significantly reduce ambiguity, provide deeper insights, facilitate theory comparison, make very precise behavioural predictions and help building automated systems. This approach also bears some implication for between-group discrimination, because participants will categorize social stimuli into in-group and out-group in each experiment. Fitting models to these type of phenomena can identify the exact cognitive mechanisms underlying discrimination between members of different groups - and highlights mechanisms involved in stereotype formation.

Beyond Formal Models, the project itself depends on creating multiple conceptual and direct replications of previous results. These experiments would form a battery of benchmark phenomena, which are essentially reliably replicated experimental findings. Replications would create the opportunity to standardize and promote the use of open-source software. Open-source software enables researchers to build transparent, editable and inspectable experiments. Open-sourced experiments allows researchers to use, scrutinize, or to modify the programs according to their needs. Once the data is collected, we test models on their efficacy of explaining ordinally different individual response patterns in the data. We will include models like error-driven attentional models, connectionist networks, adaptive clustering models, neural networks, parametric and non-parametric Bayesian models, single linear operators, drift-diffusion models.

This large-scale model comparison across multiple well-replicated social cognitive phenomena is highly novel and ambitious for both Cognitive and Social Psychology. There are some small-scale work predating the current research (Love, Meding and Gureckis's model-fitting with SUSTAIN, Andy Wills' catlearn and Klaus Oberauer's working memory benchmarks), but the project is unique due to its unprecedented scale and extensive use of Open Science techniques and Formal Modelling approaches.