Aarhus Universitets segl

CS Colloquium - Martin Møller: A Scaled Conjugate Gradient Algorithm for Fast Supervised Learning

I would like to invite you to celebrate a 25 year old high-impact paper from Department of Computer Science, AU. To my knowledge, the paper is probably the most ever cited single research result paper written by a single author, a then PhD student, from the Department of Computer Science, AU. It has been cited 3500+ times over the last 25 years, with an increasing frequency, this year to date alone 250+ citations. The paper also shows that the Department of Computer Science was an early mover in the hyped area of Machine Learning, since the topic of the paper is on algorithms for supervised learning. I have asked the author to give a talk based on the paper, and I have also asked him to talk about what his research result has been/is used for today through some examples. The talk will be followed by a little reception, as a warm up to the Katrinebjerg Christmas Lunch, taking place later the same day. See you at the talk! Kaj Grønbæk, Head of Department, Professor.

Oplysninger om arrangementet

Tidspunkt

Fredag 7. december 2018,  kl. 15:15 - 16:00

Sted

Building 5335, room 016 (Peter Bøgh Auditorium)

Original Abstract

A supervised learning algorithm (Scaled Conjugate Gradient, SCG) is introduced. The performance of SCG is benchmarked against that of the standard back propagation algorithm (BP) (Rumelhart, Hinton, & Williams, 1986), the conjugate gradient algorithm with line search (CGL) (Johansson, Dowla, & Goodman, 1990) and the one-step Broyden-Fletcher-Goldfarb-Shanno memoriless quasi-Newton algorithm (BFGS) (Battiti, 1990). SCG is fully-automated, includes no critical user-dependent parameters, and avoids a time consuming line search, which CGL and BFGS use in each iteration in order to determine an appropriate step size. Experiments show that SCG is considerably faster than BP, CGL, and BFGS.

From the 25 year old paper:

Martin Fodslette Møller (1993). A scaled conjugate gradient algorithm for fast supervised learning, Neural Networks, Volume 6, Issue 4,1993, Pages 525-533, ISSN 0893-6080,

https://doi.org/10.1016/S0893-6080(05)80056-5. (http://www.sciencedirect.com/science/article/pii/S0893608005800565)

About the speaker 

Martin Møller has a Master’s degree in Computer Science from Aarhus University in 1990, and a PhD in Computer science also from Aarhus University in 1993, under the supervision of Professor Brian Mayoh. During his PhD study he visited Carnegie Mellon University in Pittsburgh for a longer period. Martin Møller has worked with research in machine learning and decision support systems with particular focus on neural networks, optimization, and applied mathematics. After his PhD, Martin Møller started as an entrepreneur with the company IT+, which dealt with the "Advanced Java Technology”.  He sold the company to TietoEnator in 2001 and subsequently worked for three years as Senior Vice President of TietoEnator with responsibility for the banking and insurance division in Denmark. Since 2005 Martin Møller has been employed at the Alexandra Institute, starting as Research Director later as Deputy Director. His work here includes management, strategic development and implementation, as well as business development. He acts as sparring partner on research and innovation projects, matchmaking, and strategic partnerships. Finally, Martin Møller still pursue entrepreneurship as an investor and board member in several small start ups.

After the lecture everyone is invited to a cup of glögg and some traditional Danish Christmas cookies.