Please join us for a CCN Seminar with Xaq Pitkow, Assistant Professor of Electrical and Computer Engineering (Rice) and Associate Professor of Neuroscience (Baylor School of Medicine). To schedule a meeting with Xaq during his visit, please be in touch with Jessica Hauser ([email protected]).
Title: Computationally constrained control
The brain implements algorithms that choose smart actions to achieve its goals, using a tiny fraction of the power used by today’s computers on comparable tasks. Modeling these algorithms requires us to account not only for the task demands, but also for the difficulty of thinking. We call this problem setting “computationally constrained control”. I will describe two studies that incorporate computational costs into control problems. Specifically, we generalize past work on efficient coding and predictive coding, by accounting for either the representational costs of integrating information or the computational costs of performing that integration. In both cases, the predictability of the system determines phase transitions between impulsive strategies and smarter strategies, i.e. strategies when it is worth spending computational resources to achieve better performance.