~~
Speaker: Nir Admaty
M.Sc. student under the supervision of Prof. Simon Litsyn

Wednesday, June 1st, 2016 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering

Puncturing, Expurgating and Expanding the q-ary BCH Based Robust Codes

Abstract

A code that can detect errors with high probability is called a security oriented code. A security oriented code that can detect any non-zero error is called robust. A BCH based robust code is a code that is based on the columns of the check matrix of a linear BCH code. The non-binary BCH based robust code is derived from the one-error-correcting BCH code. The binary BCH based robust code is derived from the two-error-correcting BCH code. The undetected error probability Q(e) of an error e is determined by the set of codewords that mask e. The maximal error masking probability is denoted by Q ̅. A robust code is said to be optimum if there is no other code with a larger number of codewords with the same length and the same Q ̅.

The thesis presents four constructions for optimum and nearly optimum robust codes that are based on modifications of the BCH based robust codes. In particular, they generalize, puncture, expurgate, and expand the codes while preserving their robustness.

The generalized codes are a class of codes from which the BCH based code is derived. When puncturing, u redundancy symbols are deleted from the code. When expurgating, codewords which have the same value in u predefined redundancy symbol positions are grouped together and these u redundancy symbols are deleted. When expanding, a new robust code is constructed by fusing two robust codes.

The generalized and punctured codes have Q ̅ which is a power of q (the field alphabet size), and the expurgated and expanded code have Q ̅ which is not a power of q. The generalized code has a fixed code rate which is half. However, the other three constructions can increase the robust BCH based code rate while preserving its robustness.

~~ (The talk will be given in English)

Speaker:   Prof. Ron Meir
                        Department of Electrical Engineering, Technion

Monday, May 16th, 2016
15:00 - 16:00
Room 011, Kitot Bldg., Faculty of Engineering

Control Theoretic Challenges in Theoretical Neuroscience: Combining Estimation and Control

Abstract
Biology discovered feedback and control long before the first control engineers appeared on the planet, and developed sophisticated control policies operating at multiple levels from molecules to cells to organisms to populations. The vast multi-scale complexity of biological systems, the highly effective control solutions provided in biology, and the limited, yet sophisticated, toolkit available to control engineers, suggests that there is much space for interaction between biological control and engineering, where both sides stand to gain. In this talk I will focus on the interaction of sensory adaptation and control in the context of a simple partially observable sensorimotor task. I will describe an approximate analytic approach to the intractable estimation/control problem, and will show that it suggests an optimal control based explanation of observed biological phenomena. I will conclude with some of the many open problems facing the reverse engineering of biological control systems.

Short Bio: 
Ron Meir is a Professor in the department of Electrical Engineering at the Technion. He completed his PhD in Physics at the Weizmann Institute in Statistical Physics, and worked for many years in the field of Machine Learning. One of his main current interests is in control and learning in natural and artificial systems.

~~
Speaker: Gil Sadeh
M.Sc. student under the supervision of Prof. Lior Wolf and Dr. Benny Applebaum 

Wednesday, May 18th, 2016 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering

RNN Fisher Vectors for action recognition
Abstract

Recently, Recurrent Neural Networks (RNNs) have had considerable success in classifying and predicting sequences. Additionally, the Fisher Vector (FV) encoding has been widely used for pooling local features. We present the RNN-FV, which is a novel pooling method designed especially for sequential features. The methodology we use is based on FVs, where the RNNs are the generative probabilistic models, instead of the commonly used Gaussian Mixture Model (GMM), and the partial derivatives are computed using backpropagation. This proposed method is applied on sequential feature representation of videos, to achieve a new, fixed-length, discriminative video representation. We also explore different sequential feature representations of videos on which we apply our proposed method. Using our new RNN-FV based video representations, state of the art results are obtained in the task of video action recognition on two challenging datasets, UCF101 and HMDB51. We also demonstrate how to exploit the fact that the RNN is trained in an unsupervised manner in terms of the action labels, and show that training the RNN on one dataset and testing on another does not reduce performance significantly, and state-of-the-art results are achieved while using this transfer learning approach as well. We also show another surprising transfer learning result, from the task of image annotation to the task of video action recognition, which additionally improved our results.