Better Environments for Better AI - Sarah Keren
Enhancing AI Agent Capabilities through Environment Design: Promoting Multi-Agent Collaboration and Human Interaction

Traditional AI research predominantly focuses on improving AI agents' reasoning abilities to optimize their output given specific inputs. In this research, a pioneering approach is introduced, utilizing AI to design the environments in which these agents operate. By identifying the inherent capabilities and limitations of AI agents, this novel methodology seeks to modify their surroundings to maximize performance.

This talk will present various research projects with distinct design objectives, employing diverse AI methodologies to achieve optimal designs. One such project is Goal Recognition Design (GRD), which aims to adapt environments to enable observing agents to swiftly infer the goals of acting agents. Another project, Helpful Information Sharing (HIS), explores the minimal information required to disclose to partially-informed robots to ensure the achievement of their goals. Moreover, the paper demonstrates how HIS can be applied in an information market, allowing robots to exchange knowledge about the environment, facilitating effective communication and jointly maximizing their performance.

The third project, Design for Collaboration (DFC), investigates environments featuring multiple self-interested reinforcement learning agents and endeavors to promote effective collaboration among them. Throughout the discussion, the integration of these frameworks within the overarching objective of using AI to foster multi-agent collaboration and enhance interactions between robots and humans is emphasized.

This research introduces a paradigm shift by recognizing the importance of environmental design in augmenting AI agent capabilities. By leveraging AI to shape their surroundings, this approach opens new avenues for promoting collaboration among agents and facilitating improved interactions with humans. The findings from these diverse research projects contribute to a deeper understanding of the role of environment design in advancing the capabilities of AI agents, ultimately paving the way for more effective and efficient AI systems in real-world applications.

Sarah Keren is a senior lecturer (assistant professor) at the ​Taub Faculty of Computer Science at the Technion - Israel Institute of Technology. Before joining the Technion, Sarah was a postdoctoral fellow at Harvard University and Hebrew University of Jerusalem. She received her PhD from the Technion. Sarah’s research focuses on providing theoretical foundations for AI systems that are capable of effective collaboration with each other and with people. She has received a number of awards, including the ICAPS 2020 Best Dissertation Honorable Mention, the ICAPS 2014 Honorable Mention for Best Paper, the Eric and Wendy Schmidt Postdoctoral Award for Women in Mathematical and Computing Sciences, and the Weizmann Institute of Science National Postdoctoral Award for Advancing Women in Science.

Breaking the memory wall with Neural Computational Cache: packet classification and beyond - Mark Silberstein
As computer hardware evolves, processor computing capabilities scale much faster than memory performance and cache capacity.  In our work we strive to leverage this trend to design faster computer systems by replacing memory-intensive data structures with ML model inference.
As a specific example of this concept, we demonstrate a novel algorithm for range matching that turns memory-bound data structure lookups into model inference, with the help of neural networks.  Given a set of non-intersecting numerical ranges, the goal is to find the range that includes (matches) the input. Variants of this algorithm serve as building blocks in a variety of software and hardware systems, such as network switches and routers, storage controllers, DNA sequencing, and more. The key problem of conventional algorithms is that the data structure performing matching grows too large, spilling out of caches/SRAM, and constraining the number of ranges that can be handled efficiently.  Our data structure, Range Query Recursive Model Index (RQRMI), allows compressing the index by up to two orders of magnitude by representing it as a neural network model trained to predict the location of the correct range, thus replacing the index lookup with model inference.
In this talk I will show how we apply RQRMI to accelerate packet classification in a widely-used production software virtual switch Open VSwitch  by over two orders of magnitude, while performing neural network inference on a CPU in the data plane. Our recent work demonstrates that RQRMI can offer significant benefits when implemented in hardware, i.e., to scale Longest Prefix Match in network routers to support millions of rules, or to reduce the memory requirements in SSDs with a more efficient Flash Translation Layer. 
This talk is based on the published (SIGCOMM'20, NSDI'22) and ongoing work led by my PhD student Alon Rashelbach in collaboration with Prof Ori Rottenstreich.
 
Mark Silberstein is an Associate Professor at the Electrical and Computer Engineering Department in the Technion - Israel Institute of Technology. 
He is the head of the Accelerated Computer Systems Lab (ACSL)
ACSL research covers a broad range of computer systems topics, including Operating Systems and programming abstractions for emerging hardware, such as GPUs,
programmable switches and I/O accelerators; systems security, including trusted execution environments, speculative leakage in modern CPUs, and defenses against side channel attacks; and programmable networks.
Some of the projects graduated from the lab and have been adopted by industry. Mark is a recipient of several awards, including the ACM EuroSys Jochen Liedke Young Research
Award 2020, and IEEE Micro Top Picks’18 and ‘22. He regularly serves on the program committees of systems and architecture conferences including SOSP, OSDI, ASPLOS, EuroSys, MICRO, USENIX ATC, and is co-chairing EuroSys’24.
 

AI in biomedicine: Between supporting the clinicians and revealing new biology - Yoni Savir
AI has significant potential to transform medicine by improving diagnosis and treatment outcomes by facilitating decision support systems that can cope with large amounts of data, perform laborious tasks more accurately and efficiently, and identify patterns that might not be accessible to the human eye. Another promising direction is deciphering the AI representation of the studied phenomena (such as interaction network, image, or chemical structure) to gain new insights into biological and biochemical processes. In this talk, I will present a few examples of systems we study, their clinical applications, and the challenges in achieving “interpretability” of AI, particularly for medical imaging and drug discovery.
 
Yonatan Savir is an assistant professor at the Faculty of Medicine at the Technion. The research in his lab is devoted to understanding how age affects the ability of biological systems to carry out functions by taking an interdisciplinary approach using tools from experimental biology (single-cell dynamics and omics) and Engineering (AI, mathematical modeling). He received his Ph.D. in theoretical physics of complex systems from the Weizmann Institute and did his postdoctoral training as an HFSP fellow at Harvard Medical School. He has received many prestigious awards, including the Lee Segal prize for theoretical biology, the Alon fellowship, and the American Federation for Aging Research (AFAR) young investigator award. He has vast experience in harnessing artificial intelligence for health applications, co-founded a few startups, and consults for BioTech companies.
 

What are the Challenges to Analyze Big Biomedical Data? - Yael Yaniv
Despite their great promise, artificial intelligence (AI) systems have yet to become ubiquitous in the daily practice of medicine, largely due to several crucial unmet needs of healthcare practitioners. These include lack of explanations in clinically meaningful terms, handling the presence of unknown medical conditions, and transparency regarding the system's limitations, both in terms of statistical performance, as well as recognizing situations for which the system's predictions are irrelevant. Moreover, it is challenging to generate big biomedical data because it requires samples that are labeled which need human-labor. The data are generated in different formats and sometimes parts of the data are missing or include artifacts.

Yael Yaniv is an Associate Professor at the Biomedical Engineering Faculty at the Technion – Israel Institute of Technology. Before joining the Technion, Yael was a postdoctoral fellow at the National Institute of Aging, NIH, US. Yael is a leader in the field of heart rate regulation research, specifically in sinoatrial and atrial functions. Her lab has unique expertise, ranging from in vitro and in vivo cardio-physiological modelling for exploration of molecular mechanisms to feature-based and data-driven machine learning methods and data analysis, together with translation of medical needs into mobile clinical tools. She has received a number of awards, including the International Union of Physiological Sciences Travel Award, Yanai Prize for Excellence in Academic Education, Hershel Rich Technion Innovation Award and Women’s Excellence in Scientific Research Award.
 

Social Laws for Multi-Robot Coordination and a Vision for the Future of Autonomous Cars - Erez Karpas 
Agents operating in a multi-agent environment must consider not just their own actions, but also those of the other agents in the system. Artificial social systems are a well known means for coordinating a set of agents, without requiring centralized planning or online negotiation between agents. Artificial social systems enact a social law which restricts the agents from performing some actions under some circumstances. A good social law prevents the agents from interfering with each other, but does not prevent them from achieving their goals.
However, designing good social laws, or even checking whether a proposed social law is good, are hard questions.
We present approaches for verifying whether a given social law is robust under different planning formalisms, as well as techniques for automatically synthesizing a robust social law.
Inspired by this line of research, we then discuss our vision for the future of autonomous vehicles, which must share the road with human-driven vehicles in the foreseeable future. We explore the implications of having a different set of rules for the autonomous cars, and the efficiency we can gain for society as a whole by adopting this type of social law.

Erez Karpas is an Associate Professor at the Faculty of Data and Decision Sciences (formerly known as Industrial Engineering and Management) at the Technion – Israel Institute of Technology. Prior to that he was a postdoctoral fellow at the Model-based Embedded and Robotic Systems group at MIT, and before that he completed his PhD at the Technion (2012), and his M.Sc. (2005) and B.Sc. (2001) at Ben Gurion University.
His main research interest is in automated planning and robotics.