Semi-supervised Learning via Generalized MaxEnt

May 2008 - December 2009
From May 2008 until December 2009, I was a research fellow at the Max Planck Institute for Biological Cybernetics, Department Shoelkopf. As a visiting PhD student at MPI, I worked on semi-supervised algorithms for structured output prediction with Dr. Yasemin Altun.
Advised by Dr. Yasemin Altun, Max Planck Institute for Biological Cybernetics, Department Shoelkopf, Tuebingen, Germany.

Learning Probabilistic Models of Grasp Affordances

July 2009 - September 2009

In this project I worked on the problem of learning and efficiently representing discriminative probabilistic models of object-specific grasp affordances with minimal labeled grasp configurations. We train Kernel logistic regression (KLR) to map the hypothesis space of grasps into continuous class conditional probability values indicating their achievability.

We define a distance metric and an associated kernel that combine 3D position and orientation features in the object relative reference frame. While the hypothetical configurations acquired with the 3D visual model are abundant, labeled configurations are very limited as these are obtained via time-costly experiments controlled by a human observer. We propose a semi-supervised extension of KLR and a framework to combine the merits of semi-supervised and active learning approaches to tackle the scarcity of labeled grasps. Experimental evaluation shows that combining active and semi-supervised learning is favorable if there is access to an oracle, semi-supervised learning outperforms supervised learning otherwise, particularly when the labeled data is very limited.
Joint work with Oliver Kroemer, Renaud Detry, Jan Peters, Justus Piater, and Yasemin Altun.

Large Scale Manifold Transduction

May-August 2007
During summer 2007, as a research intern I worked on online large scale transduction methods at NEC Labs, Princeton.
Advised by Dr. Jason Weston, Dr. Ronan Collobert, NEC Labs Princeton, NJ

Learning Applied To Ground Robots (LAGR)

December 2004 – January 2008      (Project page @ CBLL)

LAGR is a project funded by the Defense Advanced Research Projects Agency (DARPA), which aims to promote the development of better learning systems for robot navigation in unconstrained outdoor environments. Several teams compete to get the best improvement to the baseline system developed by Carnegie Mellon University. The goal is to reach a given destination in the shortest time, using primarily two pairs of cameras as sensors, which makes obstacle avoidance in the robot’s course the main challenge.

Our lab, in collaboration with Net-Scale technologies, Inc., developed a long range obstacle detection system that allows the robot to recognize obstacles at up to 35m. In this project, I implemented a self-supervised label propagation scheme that uses location correspondences and thus enables learning by using features gathered from various views of the same obstacle, i.e., from different orientations, scales, and lighting conditions.
Advised by Prof. Yann LeCun Computational and Biological Learning Lab (CBLL), New York University.

Human Gesture Recognition

December 2002 – August 2004
Bogazici University Computer Science Department is a participant in the SIMILAR project (the European taskforce creating human-machine interfaces SIMILAR to human-human communication). The main objective of our group was to develop a human-computer interface for the disabled. I took part in the implementation of an online 3D Hand Gesture Recognition System using Hidden Markov Models, where the user controls PC applications with a colored glove. Later on, as part of my Master’s thesis, I worked on 3D Hand Posture Recognition using nonparametric optimization techniques and developed an application that distinguishes different postures of the hand such as those in sign language.
Advised by Prof. Lale Akarun, Perceptual Intelligence (PI) Lab, Bogazici University, Istanbul, TURKEY.

Cerberus: RoboCup 2002 Sony Legged Robot League Participant

September 2001 – June 2002

firefox RoboCup is an international research and education initiative that fosters robotics research. Our team, Cerberus, was a participant in the Sony Legged Robot League held in Fukuoka, Japan in 2002. The project consisted of five modules: vision, locomotion, planning and behavior, localization, and communication. As part of my undergraduate thesis, I took part in the localization module where I implemented triangulation techniques using the landmarks in the soccer field. I also developed basic machine learning tools for object detection as part of the vision module.
Advised by Prof. Levent Akin, Bogazici University, Istanbul, TURKEY.