Modeling causal learning in children

Tom Griffiths
University of California, Berkeley (UC Berkeley)

Recent work exploring causal learning in children has illustrated that children can learn causal relationships from a remarkably small amount of data -- often just a handful of observations. This presents a challenge for accounts of causal learning based on statistical inference, as standard machine learning algorithms for identifying causal relationships require significantly more data. I will suggest that the key ingredient in rapid causal learning is strong constraints on the hypotheses under consideration, outlining what these constraints might look like for learning about blicket detectors and stickball machines, and highlighting the ways in which this knowledge goes beyond representations for causal relationships such as causal graphical models.


Presentation (PDF File)

Back to Graduate Summer School: Probabilistic Models of Cognition: The Mathematics of Mind