Statistical Learning in Hybrid Systems

Statistical Learning in Hybrid Systems

The internet today allows a large fraction of the world’s population to communicate with each other with virtually no restrictions on connectivity. In contrast, in all traditional mass media (from the printing press to radio and tv), information has been technically constrained to flow through a sparse set of intermediaries (e.g., editorial rooms). Technically, these information bottlenecks usually consolidate larger quantities of information. On the internet, the flow of information is nowadays frequently structured by machine learning algorithms (e.g., recommender systems on social media) that dynamically adapt the topology of the communication graph. We propose to understand systems of humans connected via (potentially algorithmically controlled) digital networks as hybrid systems, i.e., intelligent, learning systems that combine natural and artificial intelligence. While it is plausible that large-scale deployment of such hybrid mechanisms could be having profound impact on human behavior and society, a theoretical framework for understanding potential effects is still lacking.

We propose to use mathematical concepts from statistical learning theory to model the behavior of such hybrid learning systems. Our key hypothesis is that network topology (static and adaptive) has decisive effects on how a collective is building a predictive model of the outside world from its individuals’ observations. The strengths and shortcomings of different topologies then lead to different emergent group behavior. The goal of our project is to refine this idea into a quantitative and predictive framework. This framework allows us to understand the effects of digital communication on individual and group decision-making and thus on social dynamics as well as to validate its predictions empirically in behavioral experiments. This project will focus on (1) the development of the theoretical framework, and (2) a novel behavioral experimental paradigm that tests a foundational prediction: Does the topology of the communication graph lead to different knowledge acquisition patterns and decisions in groups?

This project is funded by the Volkswagen-Stiftung