Research Projects Tim Christian Kietzmann (M.Sc.)
Institute of Cognitive Science, University of Osnabrück
Vanderbilt Vision Research Center, Vanderbilt University
Email: tkietzma (at) uos (dot) de
 
   
     
 
 

PhD Project - Invariant Object Representations across the Visual Hierachy: Neuroscientific Evidence and Computational Modeling

PhD In my PhD project, I investigate the neuronal mechanisms underlying invariant object recognition in the human visual system. In particular, I investigate potential mechanisms and features underlying viewpoint invariant object- and face-recognition, and the temporal and representational aspects of learning new object-categories. For this, I use fMRI multivariate voxel pattern analyses and EEG/MEG adaptation paradigms as experimental probes. The project is funded by a PhD position in the Neurobiopsychology lab of Professor König and by a Fulbright scholarship during my time in Professor Frank Tong's lab. phd

Current Research Projects

Viewpoint Symmetry

In this project, we investigate the cortical mechanisms underlying viewpoint invariance. Specifically, we investigate whether the human visual system exhibits selectivity for mirror-symmetric viewing angles of faces, as they are rotated in depth. For this, we use fMRI and multivariate voxel pattern analyses. Detailed information on this project can be found in the corresponding JNeuro paper.

Viewpoint Symmetry
     
Category Selectivity

In this project, we investigate the "when" and the "where" of human visual categorization behavior and categorical representations. For this, we combine extensive perceptual learning of well-controlled object categories with EEG/MEG recordings.

Category Selectivity
     
Café This project investigates whether and how natural image statistics can lead to features of increasing complexity in hierarchical systems that can be efficiently used for object recognition and the prediction of overt visual attention. cafe hierarchy

Past Research Projects

RHT

Although our understanding on conscious visual perception has made considerable progress in recent years, the role of overt visual attention in the initial perceptual formation is still controversial. There are two hypotheses that need to be considered. The first describes overt visual attention as following the perceptual formation. This theory implies that only after we consciously perceive an object's identity the visual attention is guided towards crucial features of this object. Its competitor, which assigns a more constructive role to visual attention suggests that the features that are fixated prior to the conscious perception substantially contribute to the perceptual outcome (action precedes perception hypothesis, APP). The central emphasis of this research project is to provide evidence for one of the two theories and therefore to clarify the role of visual attention in perceptual formation. The results of this project can be found in this paper.

(collaboration of NBP with Merav Ahissar at the university of Jerusalem)

FDMs of the recorded eye-movements
FILOU FILOU (Feature and Incremental Learning of Objects Utility) is a view-based object recognition system. The idea of this project is to extract key-concepts of the human visual system based on neurobiological and psychophysical findings. By translating these concepts into a suitable technical representation, a very efficient and robust object recognition system was created. It is not only capable of task-dependent feature selection, but also requires less view-prototypes to be stored than comparable approaches. This is accomplished by automatically providing each object representation with an individual amount of prototypes, depending on the complexity of the overall object and task. For more information see this and this paper. FILOU aspect graph
     
GoodGaze The goal of the GoodGaze project is to investigate how task-dependent differences in viewing behvaior can be established. By using eyetracking on webpages, two hypotheses were tested. The "weak top-down" hypothesis suggests that task-effects on fixation selection are due to relative changes of different feature channels in the bottom-up hierarchy of visual processing. Contrary to this, the "strong top-down" hypothesis sees differences in viewing behavior as being independent from changes in the bottom-up hierarchy. In this approach, the differences could for instance be established through a different spatial bias. By showing that (I) Task differences exist on overt visual attention while viewing webpages, (II) there is indeed a small interaction effect of task- and feature variables at the fixated positions, but (III) that these differences are by far not sufficient to explain the observed differences (through computational modeling), the analyses strongly favor the "strong top-down" hypothesis. More can be found in this paper. GoodGaze
NeuroRacer

The NeuroRacer is a Real-World Reinforcement Learning Application. The goal of the system is to autonomously steer the carrera car around the track in the best possible time without crashing. As part of this project, we suggested the Neuralgic Pattern Selection Algorithm (NPS). Its key idea is to concentrate sampling and batch-learning to certain "neuralgic" parts of the state space are harder to learn than others (in particular curves after long straight segments). This is achieved by calculating a failure-probability function for the track that guides later feature selection. Using this approach, it was possible to substantially reduce the amount of patterns required for successful learning while allowing for improved performance. A description of the complete system can be found here

(in collaboration with Prof. Dr. Martin Riedmiller, Machine Learning Lab, University of Freiburg)

NeuroRacer System