During everyday experience of the world, people build representations of the events they perceive, and the individuals that participate in them. In our talk, we will present a neural network model of these representations. We will focus on the representations created in Working Memory (WM): a short-term memory store whose contents reflect the results of recent experience, and encode the agent's current goals and expectations.
Our network model is founded on a particular assumption about sensorimotor processing: we assume that people experience events and individuals in the world through sensorimotor routines with well-defined *sequential structure*. Based on this assumption, we propose that events and individuals are represented in WM as *prepared sensorimotor routines*. On this view a WM representation is something 'executable', that allows an experience to be replayed, or simulated. This conception of WM has several advantages over existing models: it allows a novel implementation of the mechanism that 'binds' representations of individuals to the roles they play in episodes (e.g. AGENT or PATIENT), it supports a novel account of sentence generation, and it permits representations of probability distributions over episodes and individuals to be expressed.
Last modified: Tuesday, 03-Mar-2015 08:41:11 NZDT
This page is maintained by the seminar list administrator.