23 Sept 2016 - Chris Lucas: Seminar
Human function learning: new challenges and models
How do people learn relationships between continuous quantities? What assumptions do people rely on when interpreting continuous data? How does our past experience shape our beliefs about what relationships are better or worse explanations for noisy data? These are questions about human function learning, which is ubiquitous in everyday life, but our scientific understanding of it is still fragmentary.
In this talk, I will discuss some phenomena that current accounts of function learning cannot explain. These include results in which (1) people discover and exploit similarity between different tasks in varied and complex ways; and (2) people extrapolate confidently from sparse data in contexts where current models break down. Motivated by these results, I will introduce a new family of models based on Gaussian processes, capturing the human ability to "learn to learn" and build expertise over time, by expanding a vocabulary of abstract and composable concepts.
Time permitting, I will discuss connections to active learning, the importance of contextual information, and possible applications beyond cognitive science.