(The talk will be given in English)

Speaker:         Prof. Ofir Lindenbaum

Faculty of Engineering, Bar Ilan University

011 hall, Electrical Engineering-Kitot Building

Monday, May 15^th, 2023

15:00 - 16:00

Exploiting Randomness in Machine Learning

Abstract

Noise plays a central role in many machine learning algorithms; one well-studied example is stochastic gradient descent (SGD). Despite its simplicity, i.e., being a noisy first-order optimization method, SGD empirically outperforms gradient descent (GD) and second-order methods. In the first part of the talk, I will present Gaussian Stochastic Gates (STGs), a differentiable non-convex relaxation of the L0 norm highly effective for feature selection. Under a linear sparse regression model, I will demonstrate that our STG probabilistic feature selection formulation can recover informative features successfully. Then, I will show how a randomly aggregated least squares procedure can improve the probability of exact recovery. In the second part of the talk, I will revisit the noise model of SGD and present strong evidence that it is well characterized by a symmetric $S\alpha S$ Lévy distribution. Then, by modeling the training process as Lévy-driven stochastic differential equations (SDEs), where each parameter distributes with a different $\alpha$ value, we analyze different properties of deep nets near local minima.
(based and join works with Yutaro Tamada, Yuval Kluger, Sahand Negahban, and Barak Battash)

Short Bio

Ofir is a senior lecturer in the faculty of Engineering at Bar Ilan University. Ofir obtained his Ph.D. and M.Sc. from Tel Aviv University and his B.Sc. in Electrical Engineering and Physics (both summa cum laude) from the Technion. Following his Ph.D., he served as a Gibbs assistant professor at Yale University. His research is focused on the theory and practice of machine learning. His main goal is to enable the practical use of machine learning algorithms for scientific discovery.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בטופס הנוכחות שיועבר באולם במהלך הסמינר

You Are Invited To Attend A Guest  Lecture On Thursday, May 18, 2023 at 15:00

Room 011, Kitot Building

 By Ariel Cohen  

Title: Ultra-Fast IO and Converters Chip Design - Toward a Quarter Tera Bit-Rate Transceiver for Long-Reach Wireline Channels

:Seminar Abstract

The demand for faster, smaller, low-power and high-speed wired communication interfaces has been growing rapidly to satisfy the ever-increasing needs of datacenters and their diverse workloads. Faster electrical interfaces continue to be developed to enable this growth in datacenter bandwidth which doubles every 3-4 years. Using an ultra-fast Analog to Digital Converter (ADC) and a Digital to Analog Converter (DAC) as a part of the receiver and the transmitter, respectively, paves the way to transfer significant equalization, adaptation, and signal-processing mechanisms into the digital domain. In this lecture the motivation, the challenges, the architecture, and the circuits of such a 224 Gb/s transceiver will be presented with the focus on the receive path. The transceiver, designed in 5nm TSMC process, supports long-reach channels of 38dB attenuation at Nyquist frequency with better than 6e-7 BER. The lecture will be concluded by research opportunities of such technologies.

Index Terms—Transceiver, SERDES, Time interleaved ADC, DAC, SAR, PAM4, 224 Gb/s, digital FFE, digital DFE

Ariel Cohen Bio

Ariel Cohen received his B.Sc. degree in in electrical-engineering from Ben-Gurion University, Israel, in 1997, and his M.Ss. and Ph.D. degrees in Neuroscience from the Hebrew-University of Jerusalem, Israel, in 2003 and 2008, respectively.
Ariel is a Sr. Principal Engineer with Mixed-Signal IP group, Intel, and leads the technologies for 112/224 Gb/s ADC and DAC based SERDES since 2015. From 2008 to 2015, Ariel led the development of integrated 10 Gb/s Ethernet PHY, high-accuracy thermal sensors, sigma-delta ADCs and DACs teams.  From 2005 to 2008, he took part in establishing the bioelectronics laboratory and developed ultra-sensitive silicone-based biosensors for protein detection. Since 2012, Ariel has been an external lecturer within the Department of Electrical-Engineering and Applied Physics, The Hebrew University, Israel, teaching the VLSI courses.

His research interests include SERDES, ADCs, DACs, thermal sensors and biosensors.

Ariel Cohen was the recipient of the 2016, 2021 Intel Achievement Awards.

-סמינר זה יחשב כסמינר שמיעה לתלמידי תואר שני-

Electrical Engineering Colloquium 

Speaker: Dr. Ilai Bistritz

Title: It's Just a Game: Designing Distributed Multiagent Protocols

Abstract 

Automation relegates many decision-making processes from humans to machines. In recent years, automation is powered by machine learning, which enables trained machines to achieve expert-level performance in some tasks. Many of these tasks require humans to interact (e.g., driving, delivery, construction), so automating them will result in interacting machines where the decisions of one machine affect others. As game theory predicts, this interaction can lead to a globally inefficient equilibrium. However, machines follow programmatic objectives and protocols that, unlike humans, are not limited by selfish interests, but by information and resources. This modern paradigm calls for new tools to design efficient multiagent protocols.
 
We will first highlight the major design challenges and then discuss the distributed energy allocation problem as a concrete example. We formulate this problem as a game and study the performance of best-response dynamics (BRD) as a distributed algorithm that the sources run to allocate energy to consumers. We show that  "bad networks" exist where BRD suffers from poor performance. However, empirically, these bad networks are rare. Drawing inspiration from this empirical finding, we analyze BRD as a random process in a random game. We show that, with high probability, BRD is asymptotically optimal (in the size of the network). Applying our “random games approach” broadly may reveal that BRD is efficient more often than the worst-case overly pessimistic approach would suggest. This is an encouraging finding since BRD requires little to no coordination between the agents.

Light refreshments will be served before the lecture

This colloquium is not counted toward seminar credit.

ההרצאה לא מזכה בקרדיט שמיעת סמינרים.