Speaker: Adi Hayat

M.Sc. student under the supervision of Prof. Daniel Cohen-Or

Sunday, June 17^th, 2018 at 15:30

Room 206, Wolfson Mechanical Bldg., Faculty of Engineering

Generative Low-Shot Network Expansion​

Conventional deep learning classifiers are static in the sense that they are trained on a predefined set of classes and learning to classify a novel class typically requires re-training. In our work, we address the problem of Low-Shot network-expansion learning. We introduce a learning framework which enables expanding a pre-trained (base) deep network to classify novel classes when the number of examples for the novel classes is particularly small. We present a simple yet powerful distillation method where the base network is augmented with additional weights to classify the novel classes, while keeping the weights of the base network unchanged. We term this learning hard distillation, since we preserve the response of the network  on the old classes to be equal in both the base and the expanded network. We show that since only a small number of weights needs to be trained, the hard distillation excels for low-shot training scenarios. Furthermore, hard distillation avoids detriment to classification performance on the base classes. Finally, we show that low-shot network expansion can be done with a very small memory footprint by using a compact generative model of the base classes training data with only a negligible degradation relative to learning with the full training set.

Speaker: Itzik Nanikashvili

M.Sc. student under the supervision of Prof. Michael Margaliot

Wednesday, June 13^th 2018 at 15:30

Room 011, Kitot Bldg., Faculty of Engineering

Networks of ribosome flow models for modeling and analyzing intracellular traffic

Abstract

The ribosome flow model with input and output (RFMIO) is a deterministic dynamical system that has been used to study the flow of ribosomes during mRNA translation. The input of the RFMIO controls its initiation rate and the output represents the ribosome exit rate (and thus the protein production rate) at the 3' end of the mRNA molecule. The RFMIO and its variants can be used for studying additional intracellular processes such as transcription, transport, and more.

Here we consider networks of interconnected RFMIOs as a fundamental tool for modeling, analyzing and re-engineering the complex mechanisms of protein production. In these networks, the output of each RFMIO may be divided between several inputs of other RFMIOs. For the specific case of feed-forward network of RFMIOs we prove three important properties. First, the entire network converges to a steady-state that depends on all the translation rates in all the RFMIOs, but not on the initial conditions in the network. Second, there exists a spectral expression for the steady-state, and thus it can be determined without any numerical simulations of the dynamics. Third, the problem of dividing the output of an RFMIO between the inputs of other RFMIOs in a way that maximizes the total steady-state production rate of the network is a convex optimization problem. Hence, this problem is tractable even for very large networks. We describe the implications of these results to several fundamental biological phenomena and biotechnological objectives.

(The talk will be given in English)

Speaker:     Dr. Ittay Eyal
                        Faculty of Electrical Engineering, Technion

Monday, June 11^th, 2018
15:00 - 16:00

Room 011, Kitot Bldg., Faculty of Engineering

Teechain: Scalable Blockchain Payments using Trusted Environments

Abstract

Blockchain protocols such as Bitcoin and Ethereum exchange payments in a secure and decentralized manner, but their performance is limited due to their need to achieve consensus across a network. Cryptographic payment channels have been proposed as a second tier on top of the blockchain, allowing efficient direct payments between parties. Existing payment channel protocols, however, have two limitations: (i) Their security relies on synchronous access to the underlying blockchain, which an attacker may prevent; and (ii)~they suffer from long channel establishment times when placing collateral deposits on the blockchain.  We describe Teechain, a payment network that supports secure and scalable payments for blockchain-based cryptocurrencies using hardware trusted execution. Teechain creates chains of payment channels using multiple entangled trusted execution environments~(TEEs). Teechain is the first payment network that assumes asynchronous blockchain access and does not require bounded write-access to the blockchain. It also permits payment channels to be established near instantly by dynamically assigning deposits: deposits are not tied to a specific channel but can be moved between channels. To overcome TEE crash failures, Teechain uses a novel replication protocol between TEEs that may be of independent interest. We experimentally evaluate our Teechain implementation and show that it achieves orders of magnitude improvement in performance compared to existing solutions; with replicated Teechain nodes in a trans-atlantic deployment, we measure a throughput of between 33,000--135,000 transactions per second with 0.3 second latency, in comparison to the current state of the art which achieves 1,000 transactions per second. Moreover, Teechain payment channels place 25%--83% less transactions and data on the blockchain than existing payment solutions.

(The talk will be given in English)

Speaker:     Dr. Tamir Bendory
                        Applied and Computational Mathematics, Princeton University

Monday, June 18^th, 2018
15:00 - 16:00

Room 234, Wolfson Mechanical Bldg., Faculty of Engineering

Estimation in low SNR environment and under group action with application to cryo-EM
 

Abstract

Cryo-electron microscopy (cryo-EM) is an imaging technology that is revolutionizing structural biology, enabling reconstruction of molecules at near-atomic resolution. 
Cryo-EM produces a large number of noisy two-dimensional tomographic projection images of a molecule, taken at unknown viewing directions.
The extreme levels of noise make classical tasks in statistics and signal processing, such as alignment, detection and clustering, very challenging.
I will start the talk by studying the multi-reference alignment problem, which can be interpreted as a simplified model for cryo-EM. In multi-reference alignment, we aim to estimate multiple signals from circularly-translated, unlabeled, noisy copies.
In high noise regimes, the measurements cannot be aligned or clustered. Nonetheless, accurate and efficient estimation can be achieved via group-invariant representations (invariant polynomials). Furthermore, such estimators achieve the optimal estimation rate.
Then, I will show how this framework can be applied to the problem of 2-D classification in cryo-EM. In the last part of the talk, I will introduce the analog invariants of the cryo-EM problem and discuss how they can be used for ab initio modeling.