Predictive Modelling for Human Reasoning
Aus International Center for Computational Logic
Predictive Modelling for Human Reasoning
Vortrag von Meghna Bhadra
- Veranstaltungsort: APB 3027
- Beginn: 26. Juni 2025 um 11:00
- Ende: 26. Juni 2025 um 12:00
- Forschungsgruppe: Computational Logic
- Event series: Research Seminar Logic and AI
- iCal
AI models are often developed to solve reasoning problems optimally. In contrast, cognitive models focus on explaining and predicting replicative cognitive patterns of human information processing. And while many of the theories aim to explain an assumed ‘general’ human reasoner, only few are aimed at the individual. This talk will address the challenge of the latter by presenting the automatic generation of individualised predictive algorithms using transformer-based models. These models which have been trained on huge amounts of human data, potentially exhibit built-in cognitive patterns. Leveraging such characteristics and architecture of transformer-based models, a generalised methodology for establishing a Human-AI collaborative framework will be outlined, which can be used to generate explainable and reproducible algorithms with cross-domain applicability. While predictive accuracy and generalisability pose less of a problem, the bigger challenges in using machine learning approaches or transformer-based models may be explainability and replicability. Hence, instead of ‘just’ using such a model for directly fitting the data, it is used to extract features and to propose cognitive algorithms that are executable in systems outside of the model. Using two datasets pertaining to syllogistic and spatial reasoning, the predictive algorithms thus generated applying the presented framework, achieve mean accuracies of 68% and 81%, respectively. Both algorithms outperform other established, state-of-the-art cognitive models by far, surpassing the (previously) best state-of-the art models in syllogistic and spatial human reasoning by 19% and 13%, respectively.
