Research Topic (nur in Englisch)

---

Abstract

Natural learning processes are typically based on complex interactions between students and the teacher. Bidirectional communication - from learners to the teaching instance and vice-versa - is an essential component of learning. In contrast to this, the conventional supervised machine learning model does not incorporate communications of any kind from students to the teacher.

One question naturally arises within this context: Can a modified learning model which incorporates not unidirectional but bidirectional communication help to make machine learning more efficient? Indeed, so-called active learning, as suggest by nature, considers a learning model which employs a simple structured bidirectional communication process and helps to overcome one serious problem: As a necessary prerequisite within the conventional learning model, we need a completely labelled training set, i.e. all examples have to be assigned to the correct target objects before the actual training process starts. However, in many practical applications, labelling examples can not be performed automatically but involves human judgement or costly experimental measurements and is therefore time consuming and expensive. This disadvantage becomes even more apparent when considering the number of training examples which is necessary to learn accurate models: depending on the problem domain, this number can easily exceed several thousands.

Active learning aims at reducing the labelling effort and accelerating the learning process: Starting with only a small amount of training examples, the learning algorithm selects new training examples from a finite set of initially unlabelled examples, then requests their correct labels and incrementally learns the desired function. The key point is that by selecting only the most informative examples to be labelled, in many applications it is possible to learn a model by using less labelled examples without a significant loss of generalization accuracy in comparison to conventional learning based on the complete training set.

---

Selected Publications

• Brinker, Klaus (2004).
  Active Learning with Kernel Machines. PhD Thesis.
  University of Paderborn.
  [Download PDF] [Download BibTeX]
• Brinker, Klaus (2004).
  On Multiclass Active Learning with Support Vector Machines.
  Proceedings of the 16th Eureopean Conference on Artificial Intelligence (ECAI 2004) (pp. 969--970). Valencia, Spain.
  [Download PDF] [Download BibTeX]
• Brinker, Klaus (2004).
  Active Learning of Label Ranking Functions.
  Proceedings of the Twenty-First International Conference on Machine Learning (ICML 2004) (pp. 129--136). Banff, Alberta, Canada.
  [Download PDF] [Download BibTeX]
• Brinker, Klaus (2003).
  Incorporating Diversity in Active Learning with Support Vector Machines.
  Proceedings of the Twentieth International Conference on Machine Learning (ICML 2003) (pp. 59--66). Menlo Park, California: AAAI Press. Washington, DC, USA.
  [Download PDF] [Download BibTeX]
• Brinker, Klaus (2003).
  Active Learning of Ranking Functions.
  Proceedings of the 26th German Conference on Artificial Intelligence (KI-2003) Workshop - Preference Learning: Models, Methods, Applications (pp. 23-33). Hamburg, Germany.
• Brinker, Klaus (2003).
  Active Learning of Preferences.
  Tagungsband der GI-Informatiktage 2003 (pp. 62--65). Konradin-Verlag. Bad Schussenried, Germany.