Learning Slow Features with Reservoir Networks for Biologically-inspired Robot Localization

This work proposes a hierarchical biologicallyinspired architecture for learning sensor-based spatial representations of a robot environment in an unsupervised way. The learning is based on the fact that high-level concepts, such as the robot position, which vary in a slower timescale, can be found in a fast-varying input signal, like distance sensors. It is also assumed that the input is low-dimensional, providing limited information from the environment. The proposed architecture is composed of three layers, where the first layer, called the reservoir, is a fixed randomly generated recurrent neural network, which projects the input into a high-dimensional, dynamic space. The second layer is trained with Slow Feature Analysis (SFA), generating instantaneous slowly-varying signals from the reservoir states. Using Independent Component Analysis (ICA), the third layer implements sparse coding on the SFA output. The architecture, called RC-SFA, benefits from the short-term memory of the reservoir and the unsupervised learning mechanisms of SFA and ICA. We show that, using a limited number of noisy short-range distance sensors, mobile robots are able to learn to self-localize in simulation as well as in real environments. It is not only the current sensor reading which is needed for predicting the robot position, but also a history of the input stream. We compare the RC-SFA model with a time-delayed model using only SFA and ICA, and show that the reservoir is essential for the temporal processing of the the input stream. Results also show that the SFA and ICA layers show activation patterns which resemble, respectively, the firing of grid cells and hippocampal place cells found in the brain of rodents.