A neural dynamic model of action parsing

Humans can imitate the actions of others from early on in their life. Imitationhelps to transmit knowledge about tool usage or motor skills precisely. Precisetransmission of knowledge allows us to accumulate knowledge from past gener-ations and innovate further on top of it, which sets humans apart from otherprimates. Central to understanding how humans can learn by imitation is howhumans can extract actions from visual scenes with moving objects and buildan internal representation of them that can be utilized later.The thesis presents a neural dynamic model based on Dynamic field theory(DFT) to address how representations of actions are built based on visual input.DFT is a mathematical theory that describes the dynamics of activations ofneural populations using differential equations. Based on changes in dynamicregimes, representations of stimuli arise and cease autonomously.The model receives a video input in which an agent performs a sequence ofactions. The task of the model is to parse the video into discrete actions andgenerate a symbolic representation of each action. Descriptions of an actionsequence should be stored in memory that preserves the order in which actionsare presented.The model identifies an actor, allowing the model to understand the videoinput as a sequence of actions. Objects fulfilling certain relations with theactor are identified as playing specific roles within the description of the currentaction. Different kinds of actions can be differentiated based on which relationsare fulfilled. The model detects a point in time at which the current actionterminates, stores a symbolic representation of the action, and continues withparsing the following action. The model solves all of the sub-problems mentionedabove autonomously without interference from the modeler.By developing a neural dynamic model that can parse sequences of actions,the thesis contributes to understanding how humans can learn by imitation.