Spatio-temporal learning with the online finite and infinite echo-state Gaussian processes

Area of Science:

• Machine Learning
• Robotics
• Adaptive Systems

Background:

• Biological systems demonstrate adaptive capabilities, particularly in dynamic environments with sequential, multivariate data like robotic sensory streams.
• Existing methods struggle with noisy time-series data and irrelevant features in adaptive systems.

Purpose of the Study:

• To develop novel adaptive algorithms for learning from noisy, sequential, and multivariate data.
• To introduce the online echostate Gaussian process (OESGP) and its infinite variant (OIESGP).

Main Methods:

• The OESGP combines echo-state networks with Bayesian online learning for Gaussian processes.
• The OIESGP extends OESGP to infinite reservoirs, utilizing a novel recursive kernel with automatic relevance determination for feature weighting.
• Stochastic natural gradient descent is used for iterative kernel hyperparameter adaptation.

Main Results:

• Both OESGP and OIESGP demonstrate high accuracy on noisy benchmark problems, including one-step prediction and system identification.
• The proposed methods outperform state-of-the-art techniques and standard kernels, especially when dealing with irrelevant data dimensions.
• Insights into the underlying system dynamics can be derived from the learned hyperparameters.

Conclusions:

• The OESGP and OIESGP are effective iterative fixed-budget methods for learning from noisy time-series data.
• These algorithms show significant promise for applications in robotic learning-by-demonstration, as evidenced by case studies with the Nao robot and ARTY smart wheelchair.