Neural Support Vector Machines

Staff - Faculty of Informatics

Start date: 3 September 2012

End date: 4 September 2012

The Faculty of Informatics is pleased to announce a seminar given by Marco Wiering

DATE: Monday, September 3rd 2012
PLACE: USI Università della Svizzera italiana, room A31, Red building (Via G. Buffi 13)
TIME: 10.30

ABSTRACT:
There are many applications for machine learning algorithms. Support Vector Machines and Neural Networks are very popular algorithms, since they perform usually very well and have obtained several successes on challenging datasets. This presentation will describe a new machine learning algorithm, called the Neural Support Vector Machine (NSVM). The NSVM is inspired by SVMs and neural networks and combines the advantages of both. On one hand, in contrast to SVMs, the NSVM can cope with multiple output units, sharing the same adaptive hidden feature layer. In this way, the NSVM can also be used an an autoencoder, which is not possible for an SVM. On the other hand, the NSVM uses the objective function used in SVMs, and uses support vector coefficients to weigh the importance of training examples. This allows it to generalize better than standard multi-layer perceptrons. We tested the NSVM on classification, regression, and autoencoding problems, and the NSVM achieves state-of-the-art performance on these.

BIO:
Dr. Marco Wiering finished his Ph.D. in 1999 at the university of Amsterdam and at IDSIA in Lugano in Switzerland, where he researched novel approaches to make reinforcement learning more efficient. After that he became assistant professor in Utrecht in the Netherlands and researched different machine learning techniques between 2000 and 2007. Currently he pursues a tenure-track position for associate professor in the department of artificial intelligence at the University of Groningen, the Netherlands. He is associate editor of the IEEE Transactions on Neural Networks and Learning Systems since 2011 and he was the chair of the IEEE Computational Intelligence Society on the technical committee of Adaptive Dynamic Programming and Reinforcement Learning from 2010 to 2012. Dr. Wiering recently published the book: "Reinforcement Learning: State-of-the-Art", with Springer together with Dr. Martijn van Otterlo. Dr. Wiering has published more than 80 journal and conference papers, and has supervised or is supervising 7 Ph.D. students and 80 Master students. His research interests are reinforcement learning, machine learning, neural networks, support vector machines, computer vision, time-series prediction, computer games, multi-agent systems, big-data analysis, and robotics.

HOST: Prof. Jürgen Schmidhuber