Learning in autonomous and intelligent systems: Overview and biases from data sources

Abstract

Autonomous and Intelligent Systems (A/IS, to adhere to the terminology of the IEEE Ethically Aligned Design report) can gather their knowledge by different means and from different sources. In principle, learning algorithms are neutral; rather, it is the data they are fed during the learning period that can introduce biases or a specific ethical orientation. Human control over the learning process is more straightforward in learning from demonstration, where data sources are restricted to the choices of the demonstrator (or teacher), but even in unsupervised versions of reinforcement learning, biases are present via the definition of the reward function. In this paper we provide an overview of learning paradigms of artificial systems: supervised and unsupervised methods, with the most striking examples in each category, without too much technical detail. Furthermore, we describe the types of data sources that are presently available and in use by the robotics community. We also focus on observable bias in image datasets and originated by human annotation. We point at quite recent research on bias in social robot navigation and end with a brief reflection about ambient influences on future learning robots.