שבוע של ביקורים במעבדות המחקר, סדנאות מעשיות ושלל פעילויות מעוררות השראה

Electrical Engineering Systems Seminar

Speaker: Hagay Halperin

M.Sc. student under the supervision of Prof. Guy Even

Monday, 12^th May 2025, at 15:00

Room 011, Kitot Building, Faculty of Engineering

Class Filter - a Fast, Efficient, and Concurrent Dynamic Filter

Abstract

Filters, such as the Bloom filter, are widely used approximate membership query (AMQ) data structures. Filters are fast, compact, and have a wide range of applications. Dynamic filters, which support insertions, queries, and deletions, often have a trade-off between space efficiency and performance and are often not scalable for a concurrent use case.

In this thesis, we introduce the class filter - a dynamic filter that is space efficient, fast, robust, supports full concurrency and has scalable concurrent performance compared to other filters. The class filter’s architectures strive to minimize the mean memory access count and prioritize false query performance. We introduce the pipeline technique to increase operation performance, which may be used in other filters and dictionaries. Additionally, we obtain a space lower bound for dynamic filters with a dual-pool structure, showing that asymmetry between the pools benefits the total filter’s space efficiency. Our simulations show that the class filter structure benefits from this asymmetry.

We evaluated our implementation against many reference dynamic filters in many scenarios. Our benchmarks focus on both single-threaded and multithreaded scenarios. We also explore the ”Fixed-DB” case, where concurrent queries are executed on an instance of an immutable filter. Our implementation shows competitive performance gains in all scenarios.

הרישום לסמינר יבוצע באמצעות סריקת הברקוד למודל

Registration to the seminar will be done by scanning the barcode for the Moodle

(The talk will be given in English)

Speaker:     Prof. Aviv Tamar

                        Electrical and Computer Engineering department, Technion

Monday, May 12^th, 2025

13:00 - 14:00

Sense-Plan-Act in the age of Deep Learning

Abstract

A central paradigm in autonomous robots is the "sense-plan-act" loop, where at each time step the robot observes the environment, plans an appropriate action, executes it, and continues to the next observation. This talk focuses on the role of deep neural networks in this paradigm. I will first cover several works on learning world models for robotic manipulation, starting with the Causal InfoGAN model for rope manipulation, and more recent work based on our Deep Latent Particles (DLP) model. Here, neural networks are used to learn a model of the environment, which can be used for planning, or for directly learning a policy. The second part of the talk will focus on using neural networks to speed up a planning algorithm (e.g., tree search), based on our recent work on Bayesian Online Planning. The main idea is that neural networks can be seen as approximate posteriors in a Bayesian formulation of tree search that we propose. In this formulation, the *uncertainty* of the neural network prediction is automatically accounted for, and can be exploited for faster search.

Short Bio

Aviv Tamar is an associate professor at the Electrical and Computer Engineering department at Technion. His work focuses on reinforcement learning and robot learning. Aviv is the recipient of the Krill prize, an ERC starting grant, and best paper awards at NeurIPS and NSDI.

הרישום לסמינר יבוצע באמצעות סריקת הברקוד למודל

Registration to the seminar will be done by scanning the barcode for the Moodle