Abstract
A number of recent studies have used ideal observer models to capture human physical learning and reasoning as based on approximate mental simulation. While these approaches can match human competence in specific tasks, they are still relatively far from cognitive plausibility and are limited in their ability to capture patterns of human errors. In the current work, we train a recurrent neural network on the same physical reasoning task explored in Bramley, Gerstenberg, Tenenbaum, and Gureckis (2018) (passive condition), finding a closer match to human patterns than the ideal observer model previously used to make sense of the human judgement patterns.
Original language | English |
---|---|
Publication status | E-pub ahead of print - 1 Aug 2020 |