Deep Learning for Robot Perception and Cognition

Chapter 5: Recurrent neural networks

Avraam Tsantekidis; Nikolaos Passalis; Anastasios Tefas    Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece

Abstract

Processing temporally related data, for example, time series, using feedforward neural networks pose several challenges, for example, handling sequences of varying length. Recurrent Neural Networks (RNNs) on the other hand are able to directly handle and process such data, consisting a “natural” choice for temporal data analysis. Indeed, RNNs are able to process sequential data in a hierarchical manner, building upon an internal state, which holds a rich representation of the current and past sequence step data. In this chapter, we provide the necessary intuition ...