Professor Fabio Casati1
1University of Trento
While machine learning has made amazing progress over the last decades and perhaps even more in recent years, there are still many practical problems that fall outside its reach.
The “classical” machine learning setup consists of a process where people label data to build a “gold” dataset, then a model is trained on it and used to make predictions or take decisions.
Hybrid intelligence extends this process by bringing together human computation and machine learning in many different ways to solve a given problem, often with a tighter coupling among the two.
In this talk I will present the concept of hybrid intelligence, discuss classes of problems that can be tackled with a hybrid approach, and present different processes that achieve solutions that are efficient from a cost perspective and that meet specified quality constraints.
One of the main end goals of this research thread – yet to be achieved – is to build a meta-algorithm that, for each given problem, identifies how to best leverage and combine human and machine computations.
We will see these approaches at work on a domain likely to be of interest to any scientist, that of identifying and summarizing scientific knowledge relevant for a given research problem. In this context I will also show how a “sprinkle” of machine learning on top of human computation and analogously a sprinkle of crowdsourcing on top of ML algorithms goes a long way towards improving quality and cost.