In this talk, a fully probabilistic approach to the design and training of deep neural networkswill be presented. The framework is that of the nonparametric Bayesian learning. Both fullyconnected as well as convolutional networks (CNNs) will be discussed. The structure of thenetworks is not a-priori chosen. Adopting nonparametric priors for infinite binary matrices,such as the Indian Buffet Process (IBP), the number of weights as well as the number ofnodes or number of kernels (in CNN) are estimated via the resulting posterior distributions.The training evolves around variational Bayesian arguments.Besides the probabilistic arguments that are followed for the inference of the involvedparameters, the nonlinearities used are neither squashing functions not rectified linear units(ReLU), which are typically used in the standard networks. Instead, inspired byneuroscientific findings, the nonlinearities comprise units of probabilistically competinglinear neurons, in line with what is known as the local winner-take-all (LTWA) strategy. Ineach node, only one neuron fires to provide the output. Thus, neurons, in each node, arelaterally (same layer) related and only one “survives”; yet, this takes place in a probabilisticcontext based on an underlying distribution that relates the neurons of the respective node.Such rationale mimics closer the way that the neurons in our brain co-operate.The experiments, over a number of standard data sets, verify that highly efficient structuresare obtained in terms of number of units, weights and kernels as well as in terms of bitprecision requirements at no sacrifice to performance, compared to previously published stateof the art research. Moreover, such networks turn out to exhibit much higher resilience toattacks by adversarial examples.The presentation mainly focuses on the concepts and the rationale behind the method and lesson the mathematical details.
With image collections, both private and commercial, ever growing, efficient and effective tools for managing these repositories are becoming increasingly important. Clearly, image collections are only of use if they can be queried, yet manual annotation to enable such search is expensive, time consuming and error-prone. Luckily a lot of research in the last two decades has focused on techniques to extract useful data directly from images to facilitate searching large image repositories.
Sharif University of Technology, IR
Topic of Talk: Advances in 3D Computer Vision
Recent developments in image and video processing along with availability of powerful computers and low cost cameras have led to applicable “3D Computer Vision” techniques. It is a challenging task that requires dealing with algebraic and geometrical concepts as well as image and video processing knowledge. It employs the epipolar geometry and related optimization techniques to compute the camera internal and external parameters in order to achieve the fundamental, essential, and homography transforms among different frames that have been simultaneously captured by multiple cameras. The main applications of 3D computer vision include: 3D semantic segmentation, 3D adversarial attack and defense, point cloud classification, adversarial knowledge distillation, autonomous driving cars, virtual/augmented reality, dynamic 3D pose estimation, dynamic 3D action recognition, and so forth. In this talk, I will give a brief introduction on 3D computer vision and its various applications.
Dr. Mohammad Shokoohi-Yekta
Senior Data & Applied Scientist at Microsoft
and Instructor at Stanford University
Topic of Talk: Hottest Trends and Challenges of Deep Learning
Late-breaking developments applying deep learning in retail, financial services, healthcare, IoT, and autonomous and semi-autonomous vehicles
Why time series data is The New Big Data and how deep learning leverages this booming, fundamental source of data
What's coming next and whether deep learning is destined to replace traditional machine learning methods and render them outdated