"Perception in a variable but structured world", David Kleinschmidt (Psychology, Rutgers) -- (VIDEO RECORDING AVAILABLE)
Tuesday, April 23, 2019, 01:00am - 02:30pm
Busch Campus, Psych 101
Perception in a variable but structured world
Abstract: The relationship between observable sensory signals and states of the world is inherently variable and ambiguous. Neural systems are noisy and probabilistic, but there are also inherent ambiguities that arise from things like—in vision—the projection of a three-dimensional world onto the two-dimensional retina, occlusion, etc., and—in speech perception—coarticulation of neighboring phonemes and inherent variability in the speech production process. As such, perception is fundamentally a problem of inference under uncertainty, and can be modeled using the conceptual tools of statistical inference. My research focuses on the role that variability across contexts plays in this inference process. The relationship between observable signals and states of the world is not only probabilistic, but also changes from context to context. On the one hand, this poses challenges to perception, and successful perception thus also requires constant adaptation, which we can think of as hierarchical inference of the current generative model that probabilistically connects signals and facts about the world. On the other hand, in a structured world where contexts do not change purely randomly, the context itself provides a great deal of information about how to interpret any given stimulus, or encode stimuli for memory. However, to benefit from context in this way, you need to infer which context you are currently in and what the properties of that context are, which can be viewed as statistical inference at yet another level. I'll talk about how considering all three of these levels of statistical inference helps us understand the many strategies people use for robust speech perception, as well as extensions of these ideas to a very different domain—immediate spatial recall. Finally, I'll talk about ongoing work to quantify how much structure there is in cross-context variability in natural environments, because this provides a critical constraint on what an ideal observer can extract from the context.