Tamir Hazan, Ph.D.
Assistant professor

Faculty of Industrial Engineering and Management
Technion - Israel Institute of Technology
Bloomfield building, room 503
Technion City, Haifa 32000
tamir.hazan at technion.ac.il

[home]   [papers]   [courses]

Research interests

Attention models: Deep learning revolutionzed AI and machine learning techniques can be used to achieve human-like behavior. We explore attention models that allow to interprate and improve the prediction process of a learner.
Relevant materials: High Order Attention Models for Visual Question Answering, Factor Graph Attention, Audio-Visual Scene-Aware Dialog.

visual question answering factor graph attention scene aware dialog

Perturbation models: Statistically reasoning about complex systems involves a probability distribution over exponentially many configurations. For example, semantic labeling of an image requires to infer a discrete label for each image pixel, hence resulting in possible segmentations which are exponential in the numbers of pixels. Standard approaches such as Gibbs sampling are slow in practice and cannot be applied to many real-life problems. Our goal is to integrate optimization and sampling through extreme value statistics and to define new statistical framework for which sampling and parameter estimation in complex systems are efficient. This framework is based on measuring the stability of prediction to random changes in the potential interactions.
Relevant materials: discrete variational auto encoders, UAI 2014 tutorial, correlation clustering, online structured learning, measure concentration, learning with inverse optimization, entropy bounds and interactive annotations, sampling from the Gibbs distribution using max-solvers, learning with super-modular tasks and non-decomposable loss functions, the partition function and extreme value statistics.
max-solution perturb-max samples

Markov random fields, convex duality and message-passing: To efficiently predict outcomes in complex systems one can use graphical models and structured predictors. In this setting each predictor provides partial outcomes, e.g., the semantic labels of a region in an image, and global consistency for the structured prediction is maintained by passing messages between these regions. These concepts, emerging from Judea Pearl’s belief propagation algorithm, can be interpreted in terms of optimization theory. Applying Fenchel duality we develop the convex norm-product belief propagation, and its high-oder extensions, which enforce consistency between overlapping predictors using dual block coordinate descent. This provides us with the means to use cloud computing platforms to distribute and parallelize the prediction while maintaining consistency between its subproblems (code available, also for Amazon EC2). Estimating the parameters of region based predictors increase their accuracy in many real-life programs, and currently we achieve with these methods state-of-the-art results in various computer vision applications.