|
I am a Ph.D. student in the GRASP laboratory at the University of Pennsylvania advised by Kostas Daniilidis. I am now spending the Summer of 2019 at UC Berkeley with Sergey Levine and Chelsea Finn. I am interested in deep learning, computer vision, and robotics. Previously, I received my bachelor's degree from Czech Technical University in Prague, where I also worked on camera geometry as an undergraduate researcher advised by Tomas Pajdla. For this research, I've spent time at INRIA with Josef Sivic, and TiTech with Akihiko Torii. My name is pronounced as "Oleg". I also prefer being called that in less formal writing. Wikipedia tries to explain the pronunciation here. Google Scholar / GitHub / Email / CV / LinkedIn |
|
|
|
I am broadly interested in designing learning algorithms with properties of human intelligence, which includes problems in artificial intelligence, machine perception, and cognitive robotics. My recent interests are in temporal representation learning through generative and predictive models. Specifically, I've been working toward making machines understand phenomena like agent motion, physics, and interesting moments in time through video prediction, and make use of this undertanding for control. I am also interested in intrinsic motivation, and meta-learning. During my bachelor's, I worked on camera geometry for structure from motion and proposed an algorithm for robust estimation of camera focal length. Check out this and my other fun projects on my GitHub page. |
 
|
We propose a hierarchical predictive model that predicts a sequence starting from the high level events and progressively fills in finer and finer details. We train the model on goal-conditioned prediction on up to 80-frames (=12.5 seconds) videos.
|
 
|
We learn to infer an action representation from either motor or sensory input by using a dual variational autoencoder. By learning a dynamics model in such semi-supervised manner, we achieve both high data efficiency and planning performance.
|
 
|
We use variational inference to learn a dynamics model of image observations, and construct an agent that maximizes Bayesian surprise of the future frames. The Bayesian agent is more robust to stochastic environments than simpler prior prediction schemes.
|
 
|
We discover keyframes in videos by learning to select frames that enable prediction of the entire sequence. We show that our method improves performance of hierarchical planning by finding meaningful keyframes in demonstration data. Hover the mouse (or tap the screen) here to see the video. |
 
|
We learn to discover an agent's action space along with a dynamics model from pure video data. After a calibration stage, the model can be used to perform model predictive control, requiring orders of magnitude fewer action-annotated videos than other methods. Hover the mouse (or tap the screen) here to see the video. |
 
|
The model predicts future video frames by learning to represent the present state of a system together with a high-level transformation that is used to produce its future state. Hover the mouse (or tap the screen) here to see the video. |
|
|
|
 
|
MSE loss and its variants are commonly used for training and evaluation of future prediction. But is this the right thing to do? Hover the mouse (or tap the screen) here to see the video. |
|
|