CIS Department Talk - March 11, 2008
The Department of Computer and Information Science & The Society of
Computer Science Present
|Speaker:||Dr. J.C. Mesterharm
|Topic:||Practical On-Line Learning
|Date:||March 11th 2008, 2:30PM|
|Place:||John Mulcahy Hall, Room 138|
Supervised learning is a way to automatically assign labels to items based on the information in previously labeled items. For example, one can use information about previous atmospheric conditions and weather to predict tomorrow's weather. In this talk, we present results for a specific type of supervised learning called on-line learning.
On-line learning assumes a learning algorithm receives a stream of labeled items that it can use to continually refine the algorithm's prediction hypothesis. Soon after the algorithm makes a prediction on an item, the correct label is revealed to the algorithm. The correct label is based on a function called the target function. Surprisingly, in this model of learning, algorithms can still perform well even when an adversary is allowed to select the items from the target function.
While on-line learning has strong theoretical guarantees, our goal is to exploit some of the intrinsic advantages of on-line learning to make it more practical for realistic learning problems. This includes the following:
1) Tracking learning problems that change over time.
2) Improving the performance of certain popular on-line algorithms
when instances are generated by a distribution.
3) Removing restrictions on the learning algorithm receiving labels.
4) Expanding the types of relationships that can be learned.
Chris Mesterharm is currently a visiting professor at Fordham University. He completed his Ph.D. studies in Computer Science at Rutgers University and was previously a research assistant working on machine learning at NEC Research Institute. He earned his B.S. in Computer Engineering from Virginia Tech.
For more information, contact Ms. Danielle Aprea (718) 817-4480