Learn to Imitate Using Sequence to Sequence Network with Attention

Abstract

Imitation is an important beginning step for intelligent beings to learn more complicated concepts, especially from each other. Nevertheless, imitation by it self is arguably an ability that need to be learned. In this work, we demonstrate how simple imitation behavior can be learned by using sequence to sequence network with attention. As a proof of concept, an agent is developed that is capable of taking in an English word, letter by letter, as input and generating an output sequence of letters that make up to the same word as the input. By doing so, the agent is proved to possess the ability to imitate. The input English word is first transformed through an encoder network into a set of context vectors, which may either contain semantic information of the word or some other abstract concept meaningful only to the agent. Next, the context vectors are fed to a decoder network to generates an output sequence. It is observed that an attention mechanism in the decoder network can bring improved performance and much faster convergence during training. Imitation behavior was successfully created which in turn proved the relevance of the context vectors. Perhaps more importantly, the context vectors generated can be further processed to help with more advanced tasks such as complex concept learning and reasoning. The ideas presented in this work can be used to generate other more realistic robotic imitation behaviors as well. Tasks such as vocally repeating a spoken word/sentence heard by the agent, duplicating a sequence of moves performed by another robot, etc., can be broken down to the task of duplicating a sequence of sounds or moves codified by a sequence of letters, words, or numbers.