| Mon 27 Jul 2026 | Special colloquium: From Social Recommendation to Trustworthy AIPeople generate rich signals as they communicate, search, review, and transact. My research asks how such signals can be harnessed to connect users with the right information, products, and decisions. In the enterprise, I study how social signals can reveal informal knowledge networks in large organizations. On the web, my work explores how people express and pursue information needs through text, voice, and visual queries, and how product knowledge can be extracted and structured at marketplace scale. My current focus is on trustworthy AI for personalized information access. To illustrate this agenda, the talk will present recent work addressing the challenge of training machine learning models when ground-truth labels are delayed by weeks or months in a non-stationary, adversarial environment. This setting creates a fundamental tension between learning from fresh but incomplete feedback and relying on older, more reliable labels that may no longer reflect the current environment. I will show how we formalize these tradeoffs, develop methods to address them, and demonstrate measurable gains through both offline experiments and online A/B tests on live user traffic. | Seminar room 226 |
| Wed 8 Jul 2026 | From How to Learn to What to Learn in Multiagent Systems and RoboticsThere has been a lot of exciting recent progress on new and powerful machine learning algorithms and architectures: *how* to learn. But for autonomous agents acting in the dynamic, uncertain world, it is at least as important to be able to identify which concepts and subproblems to focus on: *what* to learn. This talk presents methods for identifying what to learn within the framework of reinforcement learning, focusing especially on applications in multiagent systems and robotics. Short Bio: Dr. Peter Stone holds the Truchard Foundation Chair in Computer Science at the University of Texas at Austin. He is Chair of the Computer Science Department, as well as Founding Director of Texas Robotics. In 2013 he was awarded the University of Texas System Regents' Outstanding Teaching Award and in 2014 he was inducted into the UT Austin Academy of Distinguished Teachers, earning him the title of University Distinguished Teaching Professor. Professor Stone's research interests in Artificial Intelligence include machine learning (especially reinforcement learning), multiagent systems, and robotics. Professor Stone received his Ph.D in Computer Science in 1998 from Carnegie Mellon University. From 1999 to 2002 he was a Senior Technical Staff Member in the Artificial Intelligence Principles Research Department at AT&T Labs - Research. He is an Alfred P. Sloan Research Fellow, Guggenheim Fellow, AAAI Fellow, IEEE Fellow, AAAS Fellow, ACM Fellow, Fulbright Scholar, and 2004 ONR Young Investigator. In 2007 he received the prestigious IJCAI Computers and Thought Award, given biannually to the top AI researcher under the age of 35, in 2016 he was awarded the ACM/SIGAI Autonomous Agents Research Award, and in 2024 he was awarded the ACM/AAAI Allen Newell Award. Professor Stone co-founded Cogitai, Inc., a startup company focused on continual learning, in 2015, and currently serves as Chief Scientist of Sony AI. | Seminar room 226 |
| Thu 25 Jun 2026 | Deep Learning and Foundation Models for BiologyDeep learning and foundation models have been used across several computational domains including text, images, audio and more. While some of the most successful examples of foundation models are in the biological domain, their use has been constrained by the availability of data, the different languages used in biology (DNA, RNA, proteins and more) and the appropriateness of model architectures for specific biological problems. In this talk I will discuss work using deep learning to study human tissues at high resolution leading to novel insights about the organization of molecules and cells. I will also present work that utilizes AI foundation models to improve the design of vaccines. Finally, I will discuss how different foundation models can be integrated to improve the ability to predict how cells respond to drugs. UPDATED BIO: Ziv Bar-Joseph, PhD is a Professor at The Mina & Everard Goodman Faculty of Life Sciences in Bar-Ilan. Prior to that he was the FORE Systems Professor of Machine Learning and Computational Biology in the School of Computer Science at Carnegie Mellon University (CMU). He is also the Co-founder and Chief Scientific Officer of GenBio AI. Dr. Bar-Joseph received his PhD in computer science from MIT in 2003. His work focuses on the analysis, modeling and integration of high throughput biological data with a specific emphasis on dynamic models. He is also developing foundation models for simulating biological data across scales. Dr. Bar-Joseph is the recipient of the Overton prize, the annual award of the Intentional Society for Computational Biology (ISCB) and the NSF CAREER award. He is also the recipient of several best papers awards including at top computational biology conferences, Recomb and ISMB. | Seminar room 226 |
| Thu 11 Jun 2026 | Multi-Robot Motion Planning: A Computational Geometry PerspectiveMulti-robot motion planning (MRMP) is a classical problem in computational geometry, in which a set of robots operate in a common workspace and must move from a given set of start positions to a given set of target positions. MRMP and its many variants have been widely investigated from both theoretical and practical perspectives, and are generally PSPACE-hard. Interestingly, however, the separation between robots plays a key role in the complexity of the problem: polynomial-time algorithms exist under some assumptions on the distance between the start and target positions. Moreover, when robots are also sufficiently separated from the obstacles in the workspace, polynomial-time algorithms with optimality guarantees can also be obtained. In this talk, I will survey recent advances in the area and present algorithms that provide trade-offs between the different separation assumptions. Based on a Joint work with Tsuri Farhana and Shalev Goldshtein. Short bio : Omrit is a faculty member at the Department of Mathematics and Computer Science at the Open University of Israel. She obtained her Ph.D. in Computer Science from Ben-Gurion University of the Negev, and was a postdoctoral researcher at Stony Brook University. Her research interest is in discrete and computational geometry, optimization problems, geometric data structures and algorithms. | Building 503 auditorium |
| Thu 4 Jun 2026 | Seeing the Unseen: From Hidden Household Items to Subsurface Soil IntelligenceWhile modern AI and vision-language models excel at analyzing visible data, they frequently struggle when tasks involve unseen or hidden targets. In this talk, we will elaborate on recent research designed to overcome these limitations across robotics, computer vision, and agriculture. Specifically, we will explore methods for equipping service robots with "commonsense" reasoning about hidden household items, utilizing hyperspectral imaging to capture signals beyond human perception for autonomous driving and fruit ripeness detection, demonstrating how to accurately classify network traffic even when packet content is completely encrypted and deploying AI-based decision support systems to predict invisible subsurface soil nutrients. Together, these projects demonstrate how leveraging indirect signals, Non-RGB modalities, and advanced modeling can unlock predictive power when the target is out of sight. BIO Dr. Oren Glickman is an Assistant Professor in the Computer Science Department at Bar-Ilan University, where his research focuses on applied AI, natural language processing, and multimodal machine learning for robotics, agriculture, and sustainability. He combines his academic work with over 20 years of industry research leadership. A principal investigator on multiple funded AgriTech and environmental projects, he holds a PhD from Bar-Ilan University, an MSc from Carnegie Mellon, and a BSc from the Hebrew University of Jerusalem. | Building 503 auditorium |
| Thu 28 May 2026 | Resilient Decision-Making for Multi-Robot Systems in the Presence of AdversariesSpeaker: Roee M. Francos, Robotics, Embedded Autonomy, and Communication Theory (REACT) Lab, School of Engineering and Applied Sciences, Harvard University This talk presents a unified perspective on resilient decision-making for multi-robot systems in environments where adversaries may influence routing, search, and detection tasks. Recently, significant progress in coordination, control, and navigation of multi-robot systems has been achieved, driven primarily by the rapid commercialization of unmanned aerial systems and drones. Yet, real-world deployment remains challenging. Most research assumes cooperative or optimally performing agents, overlooking adversarial or suboptimal behavior due to uncertainty, faults, or environmental disturbances. Resilience to malicious or malfunctioning agents remains a key limitation, as such agents can degrade efficiency and destabilize routing policies, where stability is defined as bounded cost over time. Existing stability guarantees for cooperative fleets collapse when agents deviate from plans, underscoring the need for adversarially aware planning and routing theory in safety-critical applications. I will present new theoretical and experimental results on resilient multi-agent policies for coordination, control, and learning under uncertainty and adversarial influence, focusing on routing and traffic management for fleets of aerial and ground agents for which I develop adaptive algorithms that provide provable stability, resilience, and safety guarantees. Finally, I will show how resilient cooperative search strategies address challenges posed by intelligent, coordinated external adversaries, providing provable guarantees for coverage and detection in settings such as search and rescue and pursuit-evasion. I conclude by outlining future research directions at the intersection of safety, learning, and large-scale autonomy. Bio: Roee M. Francos is currently a Computer Science Postdoctoral Fellow at the Robotics, Embedded Autonomy, and Communication Theory (REACT) Lab at Harvard University working with Prof. Stephanie Gil, focusing on development of multi-agent resilient decision-making and coordination algorithms. In 2023, he completed his PhD in Computer Science under supervision of Prof. Freddy Bruckstein, at the Technion-Israel Institute of Technology. He received the B.Sc. in Electrical and Computer Engineering from Ben-Gurion University. His research interests are in multi-agent teamwork, autonomous robotics, intelligent transportation systems, bio-inspired robotics and computer vision, focusing on collaborative algorithms for motion planning of autonomous vehicles and multi-robot learning. Roee is a recipient of the 2023 Robotics Science and Systems (RSS) Pioneers Award and the 2025 IEEE Multi-Robot & Multi-Agent Systems (MRS) Young Pioneer Award. | Building 503 auditorium |
| Tue 26 May 2026 | What Can We Learn from Neuronal Connectivity Alone?Recent advances in electron microscopy and computer vision have enabled the mapping of complete wiring diagrams, called connectomes, of brain regions and even entire brains. The emergence of these increasingly large-scale datasets has intensified the need for efficient and accurate neuronal cell type identification. Traditional approaches rely on labor-intensive analyses of molecular, anatomical, and physiological features. As a step toward fully automated neuronal cell type classification, we present NTAC (Neuronal Type Assignment from Connectivity), a method for grouping neurons into cell types based solely on synaptic connectivity. Our approach is grounded in the hypothesis that synaptic connectivity is key to determining neuronal cell types, and our results provide the strongest evidence to date supporting its validity. We further show that this connectivity-based approach extends beyond cell typing, enabling neurotransmitter classification from synaptic connectivity alone. Bio: Gregory Schwartzman is an Associate Professor at JAIST. His research focuses on large-scale graph algorithms and clustering, with recent applications to connectomics. Before joining JAIST, he was a postdoctoral researcher at the National Institute of Informatics in Tokyo and completed his Ph.D. at the Technion. | Seminar room 226 |
| Thu 7 May 2026 | The Promptware Kill Chain: How Prompt Injections Gradually Evolved Into a Multi-Step MalwareIn this talk, we examine the evolution of prompt injection attacks and show how they have gradually developed into a five-stage kill chain consisting of (1) initial access, (2) privilege escalation, (3) persistence, (4) lateral movement, and (5) actions on objectives. We begin by introducing the concept of Promptware , followed by an overview of the kill chain. We then analyze each stage in detail: the evolution of initial access (from direct to indirect prompt injection, including evasion techniques across multiple modalities); privilege escalation (from “ignore previous instructions” attacks to delayed tool invocation); persistence mechanisms (from volatile state to RAG-dependent and RAG-independent persistence); lateral movement (from none, to on-device, and ultimately off-device movement); and actions on objectives (from benign proof-of-concept messages such as “haha pwned” to full remote code execution). This talk is based on joint work, The Promptware Kill Chain , with Oleg Brodt and Bruce Schneier. | Building 503 auditorium |
| Thu 30 Apr 2026 | A Semantic Approach to Verifying Programmable NetworksAs networks become more programmable, they are increasingly built around flexible software components. While this programmability enables new functionality and faster innovation, it also makes network behavior harder to reason about. In this talk, I will present a research agenda that brings ideas from formal methods to programmable networks. In particular, I will present techniques that leverage programmable-network semantics for concurrency safety, traffic monitoring, and failure recovery. More broadly, this work illustrates how semantic foundations can help bring stronger correctness guarantees to modern networked systems. Bio Guy Amir is a Postdoctoral Researcher at Cornell University, conducting research at the intersection of formal methods, networking, and systems. He earned his Ph.D. in 2024 from the Hebrew University of Jerusalem, where he studied AI safety, focusing on formally verifying reactive AI systems and interpreting neural networks. He holds an M.Sc. in Computer Science and a B.Sc. in Computational Biology and Computer Science, both from the Hebrew University. He has received Rothschild, Fulbright, AI-Net, and Charles Clore fellowships, as well as an ICML Spotlight and KLA Award. | Building 503 auditorium |
| Thu 19 Mar 2026 | Approximately Counting and Sampling Hamiltonian Motifs in Sublinear TimeCounting small subgraphs, known as motifs, within large graphs is a fundamental challenge in graph analysis. In this talk, I will introduce a novel algorithm in the standard query model for approximately counting and almost uniformly sampling any Hamiltonian motif in sublinear time. Our approach highly simplifies prior approximate counting algorithms for edges, cliques, and stars in the standard model. This work marks a significant step toward narrowing the gap between results achievable in the standard model and the more powerful augmented query model. Joint work with Reut Levi, Dana Ron and Ronitt Rubinfeld | Zoom only ↗ |
| Thu 12 Feb 2026 | Games of ThoughtsLLMs and foundation models have been ubiquitous in our day to day lives. Yet a lot of how they take strategic decisions is still unknown and untested in long horizon scenarios such as game theoretical scenarios. This talk explores the the strategic reasoning capabilities of large language models (LLMs) in different zero sum game theoretical scenarios but individually as well as as in multi LLM agent systems, We show that by infusing LLM with cognitive hierarchy models used to characterize human thought processes can produce much stronger and less exploitable strategies Hence, emulating human decision making models can enable us to improve the reasoning capabilities of LLMs in multiagent interactions. BIO Tal Kachman is currently an assistant professor in the AI department and Donders Institute of Brain and Cognition, leading the CATALYST Lab for Complex Agent Theory, Autonomy, Learning dYnamics & Theory. He obtained his B.Sc in physics and mathematics, B.Sc in Chemistry, and M.Sc in Mechanical engineering from the Technion, and his Ph.D in physics jointly from the Technion and MIT. After graduating, he held several positions in industry: as a research staff scientist in IBM research, research engineer in AQR capital management, Quantitative researcher and later derivative trader in Optiver, a CTO and co-founder of Rhizome works, before going back into academia. | Building 503 auditorium |
| Thu 5 Feb 2026 | Intuitive and Controllable AI for 3D Geometry3D geometry powers design, perception, and simulation across various domains, including healthcare, transportation, manufacturing, and entertainment. Yet, traditional workflows for creating and editing 3D content are largely inaccessible, requiring expert knowledge and extensive manual effort. Recent advances in deep learning have enabled the generation of 3D assets through simple interfaces, such as natural language prompts. However, these AI models provide limited control, making it hard to refine the asset iteratively, edit it locally, or regulate the modification strength. As a result, the transformative potential of AI for 3D geometry remains underrealized. My work aims to bridge this gap by pioneering learning-based techniques that combine intuitive user interaction with controllable manipulation and synthesis of 3D geometry. I will show how simple yet expressive interfaces, such as point clicks and text descriptions, can drive core geometry processing and generation tasks, including interactive segmentation, localized texturing, and adjustable surface deformation. These problems pose unique challenges due to the scarcity of high-quality and diverse 3D data. I will show how to address these challenges by designing controllable 3D representations and leveraging the vast prior knowledge encompassed in powerful pretrained AI models for images and language. Looking ahead, I will outline my future research agenda, expanding the principles of intuitive and controllable AI to dynamic 3D content that interacts with its environment, and discuss how this direction can advance 3D technologies across different fields. Short Bio Itai Lang ( https://itailang.github.io/ ) is a postdoctoral researcher at the University of Chicago, working with Professor Rana Hanocka. He received his PhD from Tel Aviv University, where he was advised by Professor Shai Avidan. His research focuses on artificial intelligence for geometry processing, spanning both generative and discriminative tasks. He is particularly interested in developing innovative solutions to make 3D content creation and manipulation intuitive and controllable. | Building 503 auditorium |
| Thu 22 Jan 2026 | Fundamentals of Aligning General-Purpose AIThe field of artificial intelligence (AI) is undergoing a paradigm shift, moving from neural networks trained for narrowly defined tasks (e.g., image classification and machine translation) to general-purpose models such as ChatGPT. These models are trained at unprecedented scales to perform a wide range of tasks, from providing travel recommendations to solving Olympiad-level math problems. As they are increasingly adopted in society, a central challenge is to ensure the alignment of general-purpose models with human preferences. In this talk, I will present a series of works that reveal fundamental pitfalls in existing alignment methods. In particular, I will show that they can: (1) suffer from a flat objective landscape that hinders optimization, and (2) fail to reliably increase the likelihood of generating preferred outputs, sometimes even causing the model to generate outputs with an opposite meaning. Beyond characterizing these pitfalls, our theory provides quantitative measures for identifying when they occur, suggests preventative guidelines, and has led to the development of new data selection and alignment algorithms, validated at large scale in real-world settings. Our contributions address both efficiency challenges and safety risks that may arise in the alignment process. I will conclude with an outlook on future directions, toward building a practical theory in the age of general-purpose AI. Short bio: Noam Razin is a Postdoctoral Fellow at Princeton Language and Intelligence, Princeton University. His research focuses on the fundamentals of artificial intelligence (AI). By combining mathematical analyses with systematic experimentation, he aims to develop theories that shed light on how modern AI works, identify potential failures, and yield principled methods for improving efficiency, reliability, and performance. Noam earned his PhD in Computer Science at Tel Aviv University, where he was advised by Nadav Cohen. Prior to that, he obtained a BSc in Computer Science (summa cum laude) at The Hebrew University of Jerusalem under the Amirim honors program. For his research, Noam received several honors and awards, including the Zuckerman Postdoctoral Scholarship, the Israeli Council for Higher Education (VATAT) Postdoctoral Scholarship, the Apple Scholars in AI/ML PhD fellowship, the Tel Aviv University Center for AI and Data Science excellence fellowship, and the Deutsch Prize for PhD candidates. | Building 503 auditorium |
| Thu 15 Jan 2026 | Exploiting Computationally Bounded Adversaries in New DomainsA central insight of modern cryptography is that to guarantee the security and integrity of a task, it often suffices to defend against computationally bounded adversaries. This perspective aligns with how we model real world attackers and can enable outcomes that are provably impossible against unbounded adversaries. Although this idea has long been a driving force in cryptography and complexity theory, it has not been fully exploited in several natural and fundamental problems that have traditionally been studied in the unbounded adversary model. In this talk, we show that adopting this computational perspective enables us to surpass information theoretic limits in three classic settings: communication, randomness, and privacy. Communication: We construct efficient error correcting codes for computationally bounded adversaries that achieve information rates provably higher than those possible against unbounded adversaries. Randomness: While seedless deterministic extraction is impossible for general sources, we show that it is possible to deterministically extract and condense randomness from efficiently samplable sources. Privacy: We consider two party differential privacy and show that achieving accuracy beyond information theoretic limits is tightly connected to cryptographic power, yielding a characterization of the computational assumptions that are necessary. | Building 503 auditorium |
| Mon 12 Jan 2026 | From Learning Theory to Cryptography: Provable Guarantees for AIEnsuring that AI systems behave as intended is a central challenge in contemporary AI. This talk offers an exposition of provable mathematical guarantees for learning and security in AI systems. Starting with a classic learning-theoretic perspective on generalization guarantees, we present two results quantifying the amount of training data that is provably necessary and sufficient for learning: (1) In online learning, we show that access to unlabeled data can reduce the number of prediction mistakes quadratically, but no more than quadratically [NeurIPS23, NeurIPS25 Best Paper Runner-Up]. (2) In statistical learning, we discuss how much labeled data is actually necessary for learning—resolving a long-standing gap left open by the celebrated VC theorem [COLT23]. Provable guarantees are especially valuable in settings that require security in the face of malicious adversaries. The main part of the talk adopts a cryptographic perspective, showing how to: (1) Utilize interactive proof systems to delegate data collection and AI training tasks to an untrusted party [ITCS21, COLT23, NeurIPS25]. (2) Leverage random self-reducibility to provably remove backdoors from AI models, even when those backdoors are themselves provably undetectable [STOC25]. | Building 503 auditorium |
| Thu 8 Jan 2026 | Efficient LLM Systems: From Algorithm Design to DeploymentLarge Language Models (LLMs) have transformed what machines can do and how systems are designed to serve them. These models are both computationally and memory demanding, revealing the limits of traditional optimization methods that once sufficed for conventional systems. A central challenge in building LLM systems is improving system metrics while ensuring response quality. This talk presents approaches for reducing latency in LLM systems to support interactive applications, from scheduling algorithm design to deployment. It introduces scheduling frameworks that use lightweight predictions of request behavior to make informed decisions about prioritization and memory management across two core settings: standalone LLM inference and API-augmented LLMs that interact with external tools. Across both settings, prediction-guided scheduling delivers substantial latency reductions while remaining practical for deployment. | Building 503 auditorium |
| Wed 7 Jan 2026 | Computational and Statistical Limits in Modern Machine LearningModern machine learning systems operate in regimes that challenge classical learning-theoretic assumptions. Models are highly overparameterized, trained with simple optimization algorithms, and rely critically on how data is collected and curated. Understanding the limits of learning in these settings requires revisiting both the computational and statistical foundations of learning theory. A central question in learning theory asks which functions are tractably learnable. Classical complexity results suggest strong computational barriers, motivating a focus on “learnable subclasses” defined by properties of the target function. In this talk, I argue for a different perspective by emphasizing the role of the training distribution. Fixing the learning algorithm (e.g. stochastic gradient descent applied to neural networks), I show that allowing a “positive distribution shift”, where training data is drawn from a carefully chosen auxiliary distribution while evaluation remains on the target distribution, can render several classically hard learning problems tractable. Beyond computational considerations, I then study statistical limits of learning in modern, overparameterized models using stochastic convex optimization as a theoretical framework. While classical theory often suggests that successful generalization requires avoiding memorization, I show that memorization is in fact unavoidable: achieving high accuracy requires retaining nontrivial information about the training data and can even enable the identification of individual training examples. These results reveal fundamental privacy–accuracy tradeoffs inherent to accurate learning. Bio Idan Attias is a postdoctoral researcher at the Institute for Data, Econometrics, Algorithms, and Learning (IDEAL), working with Lev Reyzin (University of Illinois Chicago), Nati Srebro, and Avrim Blum (Toyota Technological Institute at Chicago). He obtained his Ph.D. in Computer Science under the supervision of Aryeh Kontorovich (Ben-Gurion University) and Yishay Mansour (Tel Aviv University and Google Research). His research focuses on the foundations of machine learning theory and data-driven sequential decision-making. His work has been recognized with a Best Paper Award at ICML ’24 and selection as a Rising Star in Data Science (University of California San Diego ’24). His postdoctoral research is supported by an NSF fellowship, and his Ph.D. studies were fully supported by the Israeli Council for Higher Education Scholarship for Outstanding PhD Students in Data Science. | Building 503 auditorium |
| Thu 1 Jan 2026 | High Dimensional Expanders: Structure and ApplicationsExpanders are graphs that are both edge-sparse and well connected. They have been an instrumental tool in many results in mathematics and computer science, some of which seem to have little connection to graphs at all. High dimensional expanders (HDXs) are hypergraph analogues of expander graphs — sparse and well connected hypergraphs. HDXs have already played a key role in several recent results in theoretical computer science and combinatorics. However, much of their underlying theory remains unexplored. Motivated by this, we will see what HDXs are, why they generalize expander graphs, and how they serve as a common setting for results in probability and computational complexity. I will illustrate their power by presenting new Chernoff-type inequalities for HDXs and explaining their applications to: 1. Hardness amplification - constructing functions for which the best polynomial-time algorithm performs no better than random guessing. 2. Hardness of approximation - constructing natural problems for which even approximate solutions are intractable unless P=NP. The talk will not assume prior background and is aimed at a broad audience. | Building 503 auditorium |
| Thu 25 Dec 2025 | Modern Challenges in Learning TheoryModern machine learning relies on its ability to generalize from limited data, yet a principled theoretical understanding of generalization remains incomplete. While binary classification is well understood in the classical PAC framework, even its natural extension to multiclass learning is substantially more challenging. In this talk, I will present recent progress in multiclass learning that characterizes when generalization is possible and how much data is required, resolving a long-standing open problem on extending the Vapnik–Chervonenkis (VC) dimension beyond the binary setting. I will then turn to complementary results on efficient learning via boosting. We extend boosting theory to multiclass classification, while maintaining computational and statistical efficiency even for unbounded label spaces. Lastly, I will discuss generalization in sequential learning settings, where a learner interacts with an environment over time. We introduce a new framework that subsumes classically studied settings (bandits and statistical queries) together with a combinatorial parameter that bounds the number of interactions required for learning. | Building 503 auditorium |
| Mon 22 Dec 2025 | Special colloquium: Bridging Generative Models and Visual CommunicationFrom rough sketches that spark ideas to polished designs that explain complex concepts, visual communication is central to how humans think, create, and share knowledge. Yet despite major advances in generative AI, we are still far from models that can reason and communicate through visual forms. I will present my work on bridging generative models and visual communication, focusing on three complementary domains: (1) algorithms for generating and understanding sketches, (2) systems that support exploratory visual creation beyond one-shot generation, and (3) methods for producing editable, parametric images for design applications. These domains pose unique challenges, they are inherently data-scarce and rely on representations that go beyond pixel-based images commonly used in standard models. I will show how the rich priors of vision-language models can be leveraged to address these challenges through novel optimization objectives and regularization techniques that connect their learned features with the specialized representations required for visual communication. Looking ahead, this research lays the foundation for general-purpose visual communication technologies: intelligent systems that collaborate with humans in visual domains, enhancing how we design, learn, and exchange knowledge. | Building 503 auditorium |
| Thu 11 Dec 2025 | Two Lenses on Deep Learning: Data Reconstruction and Transformer StructureDespite the remarkable success of modern deep learning, our theoretical understanding remains limited. Many fundamental questions about how these models learn, what they memorize, and what their architectures can express are still largely open. In this talk, I focus on two such questions that offer complementary perspectives on the behavior of modern networks. First, I examine how standard training procedures implicitly encode aspects of the training data in the learned parameters, enabling reconstruction across a wide range of architectures and loss functions. Second, I turn to transformers and analyze how architectural choices, such as the number of heads, rank, and depth, shape their expressive capabilities, revealing both strengths and inherent limitations of low-rank attention. Together, these perspectives highlight recurring principles that shape the behavior of deep models, bringing us closer to a theoretical framework that can explain and predict the phenomena observed in practice. | Building 503 auditorium |
| Thu 4 Dec 2025 | Fine-grained Control in Visual Content CreationGenerative models have transformed visual content creation, enabling high-quality synthesis and editing of images and videos through natural language prompts. Yet, text alone is often too ambiguous for precise control, making it difficult to specify what to edit, where changes should occur, or how much to modify. In this talk, I will present recent efforts to achieve fine-grained control in generative models, moving beyond text-only prompting toward richer interaction modalities. I will discuss three complementary directions: reference-based control for capturing identity and appearance from exemplars, spatial control for governing object placement and layout, and parametric control for defining interpretable variations such as edit strength. Together, these advances aim to bridge the gap between user intent and visual outcomes, making generative models more predictable, expressive, and intuitive tools for creation. | Building 503 auditorium |
| Thu 27 Nov 2025 | Addressing the Unexpected: Anomaly Detection and AI SafetyWhile AI models are becoming an ever-increasing part of our lives, our understanding of their behavior in unexpected situations is drifting even further out of reach. This gap poses significant risks to users, model owners, and society at large. In the first part of the talk, I will overview my research on detecting unexpected phenomena with and within deep learning models. Specifically, detecting (i) anomalous samples, (ii) unexpected model behavior, and (iii) unexpected security threats. In the second part of the talk, I will dive into my recent research on a specific type of unexpected security threat: attacks on image watermarks. I will review such attacks and present my recent work toward addressing them. I will conclude with a discussion of future research directions. | Building 503 auditorium |
| Thu 20 Nov 2025 | My Attempts to Save PolitenessTheory combination in Satisfiability Modulo Theories (SMT) deals with combining algorithms for proving formulas over integers, bit-vectors, data structures, and more. Among the most important combination methods is the polite combination method, which only requires that one of the theories admits some property, while the other theory can be any decidable theory. When polite combination was introduced, it was claimed that the required property is "politeness", a notion I will formally and intuitively describe in my talk. But later, it was shown that the correctness proof had a bug, and for the proof to work, a stronger property is needed, namely "strong politeness". This led to the following question: can politeness still be useful for theory combination? In this talk I will describe several attempts made by myself and others to answer this question positively, thus saving the politeness property. No background in theory combination or SMT is assumed, though politeness is expected. | Building 503 auditorium |
| Thu 13 Nov 2025 | Towards a White-Box Secure Fiat–Shamir TransformationThe Fiat–Shamir transformation is a fundamental cryptographic technique widely used to convert public-coin interactive protocols into non-interactive ones. This transformation is crucial in both theoretical and practical applications, particularly in the construction of succinct non-interactive arguments (SNARKs). While its security is well-established in the random oracle model, practical implementations replace the random oracle with a concrete hash function, where security is merely assumed to carry over. A growing body of work has given theoretical examples of protocols that remain secure under the Fiat–Shamir transformation in the random oracle model but become insecure when instantiated with any white-box implementation of the hash function. Recent research has shown how these attacks can be applied to natural cryptographic schemes, including real-world systems. These attacks rely on a general diagonalization technique, where the protocol exploits its access to the white-box implementation of the hash function. These attacks cast serious doubt on the security of cryptographic systems deployed in practice today, leaving their soundness uncertain. We propose a new Fiat–Shamir transformation (XFS) that aims to defend against a broad family of attacks. Our approach is designed to be practical, with minimal impact on the efficiency of the prover and verifier and on the proof length. At a high level, our transformation combines the standard Fiat–Shamir technique with a new type of proof-of-work that we construct. We hope that our transformation will help preserve the security of systems relying on the Fiat–Shamir transformation. Joint work with Gal Arnon. | Seminar room 226 |
| Thu 6 Nov 2025 | No colloquium |
| Thu 23 Oct 2025 | No colloquium |
Prof. Gal Kaminka named AAAI Fellow
