Statistical Learning Theory Day

Where: Faculty of Computer Science HSE (3 Kochnovsky Proezd), room 622

Victor Lempitsky (Skoltech)

Creative Deep learning

Originally, deep learning methods were designed to recognize things, e.g. to classify images, and the first successes of deep learning were all about recognition. In recent years, however, deep learning has been used more and more often to create and to transform things (e.g. images) rather than to just recognize them. In the talk, I will discuss some of the recent projects at Skoltech Computer Vision group that develop deep architectures for image processing and generation, as well as some projects that investigate how the generated images can be used creatively. 

Stamatios Lefkimmiatis (Skoltech)

Universal Denoising Networks: A Novel CNN-based Network Architecture for Image Denoising

In this talk I will present a novel deep network architecture for learning discriminative image models that are employed to efficiently tackle the problem of grayscale and color image denoising. The motivation for the overall design of the proposed network stems from modern variational methods that exploit two basic image properties, namely the local image regularity and the non-local self-similarity. Based on this idea, I will introduce two different variants. The first network involves convolutional layers as a core component, while the second one relies instead on non-local filtering layers. As opposed to most of the existing neural networks, which require the training of a specific model for each considered noise level, the proposed networks are able to handle a wide range of different noise levels using a single set of learned parameters. Moreover, they are very robust when the noise degrading the latent image does not match the statistics of the one used during training. Extensive experiments, comparing several state-of-the-art methods, show that the proposed networks achieve state-of-the-art result, while they depend on a more shallow architecture with the number of learned parameters being almost one order of magnitude smaller than competing networks. Finally, I will highlight a direct link of the proposed non-local models to convolutional neural networks. This connection is of significant importance since it allows us to take full advantage of the latest advances on GPU computing in deep learning and makes our non-local models amenable to efficient implementations through their inherent parallelism.

Gonzalo Ferrer (Skoltech)

Robot Navigation in Dynamic Environments

The task of navigating, that is, moving from one place to another in any kind of environment, is an extremely easy task for humans. Robots on the other hand, barely perceive the world, which is formidably complex and process this limited data to plan their motions. One can argue that on simple scenarios, the task of navigating is completely solved. Nonetheless, full autonomy in robotics has not arrived yet. This is a key aspect for the future deployment of robots in order to be a mainstream technology adopted by society, either if robots are mobile platforms, autonomous cars, flying quadcopter, etc.

In this talk, I will present an overview of my work on robot navigation on dynamic environments. Under the interaction with pedestrians, complex situations arise where known path planning techniques provide poor solutions. I will present a new prediction approach on human motion and how to integrate it under the same planning scheme, obtaining a more intelligent robot motion behavior.
Still, some degree of uncertainty is unavoidable, due to the unpredictable nature of pedestrians, making impossible a perfect accuracy on prediction. Hence, I will discuss on how to calculate plans on adversarial scenarios, leveraged by probability distributions, as an effective way to avoid potentially dangerous situations. 

Приглашаем всех сотрудников, студентов и аспирантов факультета компьютерных наук.