Challenges in Interactive Machine Learning
- PDF / 555,655 Bytes
- 4 Pages / 595.276 x 790.866 pts Page_size
- 85 Downloads / 304 Views
DITORIAL
Challenges in Interactive Machine Learning Toward Combining Learning, Teaching, and Understanding Stefano Teso1 · Oliver Hinz2
© Gesellschaft für Informatik e.V. and Springer-Verlag GmbH Germany, part of Springer Nature 2020
1 Why Interactive Machine Learning?
2 About this Special Issue
The recent success of AI, and of machine learning in particular [8], has unlocked a number of high-impact applications, from machine translation to medical diagnosis, to image and video generation, to game playing. The bulk of these success stories, however, take place in relatively “lab” settings [3, 11]. With this we mean settings in which the cost of failure is limited and mistakes are easily spotted, large repositories of data are available and more supervision can be acquired relatively cheaply, noise patterns are somewhat predictable, the rules of the game and the knowledge they rely on are crisply defined, and the objective function is fully specified from the get-go [6, 10]. Most importantly, many of these success stories take place in settings — like passive learning — that are completely devoid of people [1].1 This is not insignificant, because once humans are added to the equation, all of the simplifying assumptions listed above break down. As AIs become ever more ubiquitous, the question becomes: will they work appropriately once deployed in the real world, in which humans feature so prominently? It seems prudent to prepare for a negative answer by investigating appropriate human-aware approaches.
This special issue is dedicated to interactive machine learning, in which the goal is precisely to design adaptive agents that support meaningful and beneficial interaction with humans. The article by Nadj et al. identifies and discusses design principles for interactive labeling systems by conducting a literature review. We believe that this contribution can represent a helpful starting point for further efforts to refine and expand the design of interactive labeling systems. The history of interactive learning is rooted in query learning [2]. Here, the structure of the interaction is fixed: a student model, usually a classifier, acquires supervision by asking questions to one or more human teachers, usually domain experts. This helps explaining why the most studied — and thus the most well-known — issues in interactive learning are related to acquiring supervision, e.g., balancing between the cost of annotations and the amount of information they carry, as in active learning [14], or measuring and improving the quality of annotations and annotators, as in crowd-sourcing applications [5]. The range of issues spanned by interactive machine learning, however, is even broader. The articles in this issue constitute a representative and diverse sampler. Including humans in the process, certainly comes with higher costs for manual labor. Therefore it is imperative to offer efficient processes and systems that minimize this part of the entire interactive machine learning process. Baur et al. make a significant contribu
Data Loading...
