Alon Faktor

Director of AI researchVimeo

Bio:

Alon Faktor is a Director of AI research at Vimeo, the world’s largest private video hosting platform. Alon works on innovative applications for video consumption and interaction aiming to make video content dynamically used to increase value for Vimeo customers. Alon’s research focuses on efficient multi-modal large language models and video-RAG techniques for video understanding and indexing. Alon holds a B.Sc. in Physics and Electrical Engineering from the Technion and a M.Sc. and PhD in Computer Science from the Weizmann Institute of Science.

Title:

From Passive Viewing to Active Dialogue: Redefining Video Consumption

Abstract:

At this lecture, we will present how Vimeo is harnessing the latest AI techniques to enable novel video viewing experiences. We will give a deep dive into our talk-to-video technology which utilizes RAG and LLMs on the video domain. We will demonstrate our approach to extract and index the multimodal video information such that it can be effectively utilized with the RAG architecture. We will also present several applications which we have built on top of this method such as video Q&A and library search.

Matan Levy

PhD candidate Hebrew University of Jerusalem

Bio:

A Computer Science Ph.D. candidate at the School of Computer Science and Engineering at the Hebrew University of Jerusalem, under the joint supervision of Prof. Dani Lischinski and Dr. Rami Ben-Ari

Title:

Task-Specific Adaptation with Restricted Model Access

Abstract:

The emergence of foundational models has greatly improved performance across various downstream tasks, with fine-tuning often yielding even better results. However, existing fine-tuning approaches typically require access to model weights and layers, leading to challenges such as managing multiple model copies or inference pipelines, inefficiencies in edge device optimization, and concerns over proprietary rights, privacy, and exposure to unsafe model variants. In this paper, we address these challenges by exploring ``Gray-box'' fine-tuning approaches, where the model's architecture and weights remain hidden, allowing only gradient propagation. We introduce a novel yet simple and effective framework that adapts to new tasks using two lightweight learnable modules at the model's input and output. Additionally, we present a less restrictive variant that offers more entry points into the model, balancing performance with model exposure. We evaluate our approaches across several backbones on benchmarks such as text-image alignment, text-video alignment, and sketch-image alignment. Results show that our Gray-box approaches are competitive with full-access fine-tuning methods, despite having limited access to the model.

Ofer Rosenberg

AI System/Software ArchitectQualcomm

Bio:

Since 2008, Ofer has been involved in Compute and Heterogeneous environments, contributing to the OpenCL Specification while working on CPU-based Compte and GPGPU. In 2016, he began focusing on AI at Qualcomm, when it was still called “Machine Learning”. He initially worked on the Snapdragon Neural Processing Engine (SNPE) before becoming the Software Architect for Qualcomm’s Cloud AI100. Over the past year, Ofer has concentrated on enabling Generative AI across the stack, from conversion tools to the newly created Genie (Generative AI Inference Engine) add-on to QAIRT SDK.

Title:

Generative AI on Mobile Devices: Challenges and Innovations

Abstract:

The session explores the advancements and challenges of implementing Generative AI on mobile devices. It starts by emphasizing the benefits of running Generative AI on mobile platforms, including enhanced privacy, personalization, performance, and offline support. It then describes the characteristics of a typical mobile platform, such as the Snapdragon 8 Elite, and the limitations it introduces, such as memory bandwidth and power/performance tradeoffs. The session highlights innovative solutions that enhance the performance and accuracy of models running on mobile devices, including quantization techniques, speculative decoding, and Low-Rank Adaptation (LoRA).

Danny Astmon

Founder & CEOCognata

Bio:

Danny is an expert in ADAS/Autonomous cars and deep learning, with a record of more than 20 years. Served as Harman’s Director of ADAS/Autonomous car and senior director of machine learning. He has co-founded two start-up companies, Picitup, and iOnRoad. Danny holds several United States utility patents and has created a pipeline of dozens of patent-pending applications. Danny is a graduate of the prestigious Israeli Defense Forces (IDF) Haman Talpiot program, where he served in the elite Unit 8200, and holds a B.Sc. degree in Physics from Tel-Aviv University.

Title:

Supervised Generative AI – Real Data Augmentation with DriveMatriX

Abstract:

Cognata introduces DriveMatriX, an advanced supervised generative AI platform designed to transform real-world images and videos into diverse training and validation datasets. By leveraging a foundation of real-world data, DriveMatriX maximizes coverage through controlled, repeatable, and predictable augmentations. The platform simulates complex environmental conditions—such as adverse weather, varying illumination, and dynamic urban landscapes—ensuring robust AI model performance. Integrated validation mechanisms streamline AI training workflows, improving precision, recall, and edge-case performance. DriveMatriX sets a new benchmark in generative AI by combining realism with control, accelerating development cycles while maintaining dataset fidelity and consistency.

Tamar Kashti

AI Hub Research Manager Weizmann Institute of Science

Bio:

Tamar Kashti is a seasoned leader in algorithms and AI with over 15 years of research experience across academia and industry. She specializes in deep learning, computer vision, and image processing. Tamar currently manages the AI Hub at the Weizmann Institute, where she oversees the AI internship program for students and research fellows and advances cutting-edge medical imaging projects. Dr. Kashti has led algorithmic teams at Landa Labs and HP Indigo, pioneering innovations in calibration and print technologies, earning 7 patents. She has authored 15 published papers and holds a Ph.D. in theoretical physics from the Weizmann Institute.

Title:

Deep Denoising of Multiplexed Mass-based Images: Supervised vs. Self-supervised

Abstract:

Multiplexed mass-based imaging (MBI) technologies offer transformative insights into cellular diversity but are often hindered by significant noise. This study presents a deep learning-based approach to denoise MBI data, comparing supervised and self-supervised methods. Both approaches effectively address noise-related artifacts, with the supervised method excelling after fine-tuning for specific datasets, while the self-supervised approach demonstrates strong generalization across diverse data. These methods significantly reduce manual effort, delivering high-quality, denoised images within minutes. By enhancing image usability and analysis efficiency, this automated solution accelerates the adoption of MBI technologies for researchers and clinicians in biological and clinical studies.

Ron Shapira Weber

Ph.D. Student Ben Gurion University

Bio:

Ron Shapira Weber is a Ph.D. student at Ben Gurion University (BGU) at the Vision, Inference, and Learning (VIL) under the supervision of Dr. Oren Freifeld at the Computer Science Dept. His interest areas include time series analysis and computer vision, with applications to time series joint alignment and averaging, image registration, and video analysis. He did his master's in Cognitive Science at BGU as a part of VIL group, under Dr. Oren Freifeld, and of the Computational Psychiatry Lab under Dr. Oren Shriki. Between 2019-2021 he worked as an algorithm researcher at BeyondMinds.

Title:

SpaceJAM: a Lightweight and Regularization-free Method for Fast Joint Alignment of Images

Abstract:

Unsupervised Joint Alignment (JA) of images faces high complexity, geometric distortions, and convergence to poor optima. Vision Transformers provide powerful features but do not fully resolve these challenges, leading to reliance on expensive models with heavy regularization and extensive hyperparameter tuning. We propose the Spatial Joint Alignment Model (SpaceJAM), a compact architecture with ~16K trainable parameters requiring no regularization. Evaluations on SPair-71K and CUB demonstrate that SpaceJAM matches state-of-the-art performance while offering at least a 10x speedup. By setting a new standard for rapid, effective image alignment, SpaceJAM makes JA more accessible and efficient. Code is available at: https://bgu-cs-vil.github.io/SpaceJAM/

Rotem Benisty

MSc student Technion - Israel Institute of Technology

Bio:

Rotem is an MSc student in Electrical and Computer Engineering at the Technion, where she is jointly mentored by Dr. Moti Freiman from the Faculty of Biomedical Engineering and Prof. Moshe Porat from the Faculty of Electrical and Computer Engineering. She holds a B.Sc. in Electrical and Computer Engineering from the Technion. Rotem’s research focuses on advancing isotropic MRI restoration from anisotropic data through innovative generative multi-plane deep learning techniques. Alongside her research, Rotem works as an engineer in the networking department at Nvidia, bringing several years of practical industry experience.

Title:

SIMPLE: Simultaneous Multi-Plane Self-Supervised Learning for Isotropic MRI Restoration from Anisotropic Data

Abstract:

Magnetic resonance imaging (MRI) is essential for medical diagnostics but often produces anisotropic data with varying resolutions, limiting volumetric analysis and diagnostic precision. Current super-resolution methods primarily interpolate missing slices, relying on indirect mappings and limited isotropic data, without fully leveraging MRI's inherent 3D structure. We introduce "SIMPLE," a Simultaneous Multi-Plane Self-Supervised Learning approach for isotropic MRI restoration from anisotropic data. SIMPLE utilizes multi-plane clinical scans to enhance slice quality while addressing 3D structural information to generate realistic isotropic volumes. Experiments on brain and abdominal datasets demonstrate SIMPLE’s superiority over state-of-the-art methods, improving volumetric analysis, 3D reconstructions, and clinical diagnostics.

Idan Abudi

Algorithm Researcher 4M Analytics

Bio:

Idan Abudi is an Algorithm Researcher at 4M Analytics, a company specializing in providing cutting-edge AI solutions for mapping underground infrastructures on a large scale. Before joining 4M Analytics, Idan served as a researcher and commander at Unit 9900 for almost 9 years (mainly focusing on Geospatial Intelligence, with expertise in geographic and visual data research). He holds an MSc in Financial Math from Bar Ilan University.

Title:

Utilizing aerial imagery for utility object detection in infrastructure mapping

Abstract:

Underground infrastructure includes critical utility systems like water, gas, electricity, and communication networks, as well as transportation tunnels and structural foundations. Accurate mapping of these hidden systems is essential to prevent costly damages and ensure effective project planning. At 4M Analytics, we introduce a groundbreaking solution that harnesses aerial imagery from planes, integrating state-of-the-art object detection with domain-specific innovations. Our innovative approach addresses significant challenges, such as object variability and resolution limitations.

Yael Sde-Chen

Algorithm DeveloperApplied Materials

Bio:

Yael is an Algorithm Developer at Applied Materials, driving cutting-edge solutions in the semiconductor process control field. She holds a M.Sc. in Electrical Engineering from Technion. Yael has 5 years of industry experience from Applied Materials, Amazon and the IDF, with focus on deep learning methods for semantic segmentation, data augmentation, and generative models.

Title:

Repair Blind Spots in Semantic Segmentation

Abstract:

Semantic segmentation of images is crucial in AI expert systems. The Unet network architecture is widely considered the most practical choice for CNN solutions in application-specific semantic segmentation.

Despite their success, the Unet architecture has inherent problems that prevent true shift-equivariance and single pixel-level accuracy, as we demonstrate below.

Here we propose alternative CNN architecture, offering higher accuracy and efficiency, which is specifically designed for mission-critical applications where each pixel is crucial, such as anomaly detection in medical imaging, industrial quality control, and more.

Rajaee Khateb

PhD Student Tel Aviv University

Bio:

I am an Electrical Engineering Ph.D. student at Tel-Aviv University, advised by Prof. Raja Giryes.

My research is in the area of Artificial Intelligence and it focuses on the usage of 3D data. My research goal is to explore new 3D representations, and their integration with Generative AI methods such as Diffusion Models. Our vision is to use this unique integration in order to take the current 3D solutions a step forward.

Previously, I obtained a B.Sc. (Cum Laude) and M.Sc. from the department of Computer Science at the Technion, where I was advised by Prof. Michael Elad. My M.Sc. thesis was about unfolding greedy sparse pursuit algorithms into deep neural networks.

Title:

TriNeRFLet: A Wavelet Based Triplane NeRF Representation

Abstract:

In recent years, the neural radiance field (NeRF) model has gained popularity due to its ability to recover complex 3D scenes. Following its success, many approaches proposed different NeRF representations in order to further improve both runtime and performance. One such example is Triplane, in which NeRF is represented using three 2D feature planes. This enables easily using existing 2D neural networks in this framework, e.g., to generate the three planes. Despite its advantage, the triplane representation lagged behind in 3D recovery quality compared to NeRF solutions. In this work, we propose the TriNeRFLet framework, where we learn the wavelet representation of the triplane and regularize it. This approach has multiple advantages: (i) it allows information sharing across scales and regularization of high frequencies; (ii) it facilitates performing learning in a multi-scale fashion; and (iii) it provides a `natural' framework for performing NeRF super-resolution (SR), such that the low-resolution wavelet coefficients are computed from the provided low-resolution multi-view images and the high frequencies are acquired under the guidance of a pre-trained 2D diffusion model. We show the SR approach's advantage on both Blender and LLFF datasets.

Bella Specktor Fadida

Lecturer University of Haifa

Bio:

Bella leads the Healthcare Computer Vision (hVision) lab in the Medical Imaging Sciences department at the University of Haifa. She earned her PhD from the Hebrew University, specializing in medical imaging, and previously spent seven years developing medical imaging algorithms at Philips. Bella serves as a board member of Women in MICCAI (WiM) and is the founder and organizer of the Machine Learning for Medical Imaging (MLMI) and Haifa Machine Learning meetups.

Title:

Fetal weight estimation using deep learning-based segmentation

Abstract:

This work presents a deep-learning method for whole-body fetal segmentation from MRI and evaluates its repeatability, reproducibility, and accuracy. A normal MRI-based fetal weight growth chart was created for the first time, and sensitivity in detecting fetal growth restriction (FGR) was assessed. Retrospective data from 348 fetuses were analyzed, with assessments for repeatability (n=22), reproducibility (n=6), and accuracy (n=7). The model achieved high segmentation accuracy (Dice=0.973), strong agreement with the gold standard, and reliable fetal weight estimation. The MRI-based growth chart was consistent with ultrasound charts and accurately identified FGR cases, demonstrating its potential for fetal growth assessment.

Georgy Melamed

Principal ScientistWalmart Global Tech

Bio:

Georgy Melamed is a principal scientist at Walmart Global Tech Israel. He has over 20 years of experience in CVML domains at military industry, startups and corporate. He holds an M.Sc. in Electrical Engineering from Tel Aviv University and is an alumni of Talpiot program.

Title:

Generative video virtual try-on

Abstract:

Visualizations are critical in e-commerce, both for marketing and customer-vendor expectations matching. The particular use case of garment virtual try-on underwent a tremendous growth in the past decade, accelerating exponentially with the present leap of generative models. The specific differentiator between an impressive presentation and real-life solution is trustworthiness of the result, mitigating customer's risk during remote purchase.

Video is the next frontier for virtual try-on. In this talk we present an evolution of Zeekit's historic SOTA static solution, acquired by Walmart several years ago, into a contemporary video try-on feature, leveraging recent technology.

Sharon Peled

M.Sc. student Technion - Israel Institute of Technology

Bio:

Sharon is an M.Sc. student at the Technion's Data and Decisions Faculty, co-advised by Dr. Moti Freiman and Dr. Yosi Maruvka.

His research focuses on developing Multiple Instance Learning (MIL) models for gigapixel histopathological images (Whole Slide Images), with an emphasis on creating scalable solutions for clinical diagnostics.

In addition to his academic work, Sharon has several years of experience in signal processing and machine learning research and currently serves as a senior lead researcher at the IDF.

Title:

PSA-MIL: A Probabilistic Spatial Attention-Based Multiple Instance Learning for Whole Slide Image Classification

Abstract:

Whole slide images are gigapixel-sized digital scans of tissue samples, designed to capture intricate cellular and morphological patterns.

Their immense size necessitates division into smaller tiles, typically analyzed as instances within a Multiple Instance Learning (MIL) framework.

However, this formulation often overlooks spatial relationships between instances, which are crucial for capturing complex tissue structures during a histopathological examination.

To this end, we present a novel attention-based MIL framework that utilizes a probabilistic interpretation of self-attention to dynamically infer spatial dependencies during training.

Our approach eliminates the need for predefined spatial assumptions, enabling flexible spatial modeling and achieving state-of-the-art performance.

Rami Ben-Ari

Principal Research Scientist OriginAI

Bio:

Rami Ben-Ari is a Principal Research Scientist and Technical Lead at OriginAI and an Adjunct Professor at Bar-Ilan University in CS-EE Faculty. Actively engaged in the academic community, he co-supervises MSc and PhD students and has authored over 50 papers, along with numerous patents. His research focuses on deep learning techniques for image retrieval, multimodal learning, video understanding, and generative models. He holds a PhD in Applied Mathematics from Tel-Aviv University, specializing in computer vision.

Title:

Fast image inversion and editing with diffusion models

Abstract:

Emerging diffusion models have demonstrated impressive capabilities in generating images from textual prompts and sampled random noise, commonly referred to as a seed. A common approach to edit a real image involves adjusting the prompt while evaluating its corresponding seed, a process known as image inversion.

In this talk, I will introduce a novel and efficient solution for image inversion and editing with diffusion and flow matching models, using the well-known Newton-Raphson numerical scheme. Our method, Guided Newton-Raphson Inversion (GNRI), enables high-quality reconstructions and edits, outperforming existing inversion and editing techniques in both accuracy and speed. Notably, GNRI requires no model training, fine-tuning, prompt optimization, or additional parameters. This work has been accepted for publication at ICLR 2025.

Ben Vardi

PhD studentReichman University

Bio:

Ben is a Computer Science PhD student at Reichman University, supervised by Professor Ariel Shamir. His research focuses on the robustness and limitations of vision-language models. Previously, he worked as a computer vision engineer at Snap. He holds an MSc in Computer Science from Ben-Gurion University and a BSc in Biology and Psychology with an emphasis on Neuroscience from Tel Aviv University.

Title:

CLIP-UP: CLIP-Based Unanswerable Problem Detection for Visual Question Answering

Abstract:

Recent Vision-Language Models (VLMs) excel in tasks like multiple-choice Visual Question Answering (VQA) but often struggle with unanswerable questions, providing incorrect responses to irrelevant queries. We propose CLIP-UP, a lightweight method leveraging CLIP-based alignment to equip VLMs with the ability to detect and withhold answers to unanswerable multiple-choice VQA. By training a few additional layers while preserving the original weights, CLIP-UP applied to LLaVA models achieves state-of-the-art performance on the MM-UPD benchmark for unanswerable question detection.

Nizan Mashall

M.Sc studentTechnion - Israel Institute of Technology

Bio:

Nizan Mashall is a mechanical engineer and a master’s student in the Technion’s Autonomous Systems and Robotics program. He previously worked at Rafael and BeautAI. Under the joint supervision of Prof. Miriam Zacksenhouse and Prof. Erez Karpas, Nizan develops artificial intelligence for autonomous robotic capabilities. His current work on affordance grounding in 3D space was accepted to the AAAI Bridge Program for Foundation Models and Robotics (PLAN_FM).

Title:

From Semantic Understanding to Geometric Features: Using Foundation Models for Novel Robotic Tasks

Abstract:

Foundation Models possess implicit knowledge of objects and their use. Our system leverages this to detect geometric features (vertices, edges, planes) and define task-specific coordinate systems, enabling model-based planners to execute manipulation tasks without task-specific training. Unlike current vision-language approaches requiring demonstrations or fine-tuning, our fully automated pipeline integrates five pre-trained foundation models. It generates digital twins, identifies key geometric features, and transfers these detections to real-world object point clouds. This approach achieves robust zero-shot performance across diverse objects and tasks, advancing adaptable robotic systems for autonomous manipulation in dynamic environments.

Oran Shayer

AI Research LeadAppsFlyer

Bio:

Oran has over a decade of experience in machine learning, computer vision, and AI. He has a track record in both pure and applied research, leading AI research groups and managing end-to-end research to production processes. Prior to AppsFlyer, Oran worked at startups as well as large corporations (Apple, GM, Intel). Oran holds a B.Sc. and M.Sc. in electrical engineering from the Technion.

Title:

Designing VLM-Based AI Agents for Large-Scale Video Analysis

Abstract:

Dive deep into innovative architectures for building vision language models (VLMs)-based AI agents for large-scale video processing pipelines. By reflecting on key aspects of VLMs' reasoning and processing mechanisms, we propose several agent architectures designed to fully leverage their capabilities. Through practical examples, we'll explore strategies to maximize VLMs' visual understanding capabilities and design production-grade performance VLM-centric agents, while also suggesting more cost-effective alternatives.

Ben Fishman

Ph.D Candidate Bar Ilan University

Bio:

Ben Fishman is an AI & Algorithms researcher and manager with 11 years of experience in both industry and academia. Today he is a Computer Science Ph.D. candidate at Bar Ilan University, supervised by Prof. Gal Chechik and Dr. Idan Schwartz focusing on Generative AI & Multi Modal. Prior to that he served as a Director of AI & Algorithms at Microsoft and as a team leader at Mobileye.

Ben’s main areas of interest include Computer Vision, Speech & Audio, Deep Learning, and LLMs. He holds an M.Sc. in Electrical Engineering and a B.Sc. in Biomedical Engineering both from Tel Aviv University.

Title:

Bringing AI to Production

Abstract:

Today 85%-90% of AI projects fail to become a real-world application. In this session, we will explore the end-to-end lifecycle of an AI project, focusing on the challenges encountered at each stage and how to address them.

Reaching production involves much more than training a deep learning model and therefore we will highlight best practices in other crucial areas: Data management (collection, annotation, curation, and engineering), infrastructure (MLOps, DevOps, tools) for an efficient and smooth development process, and algo-evaluation for an effective process with insights for the DS team.

The session will provide actionable insights to help AI projects succeed in delivering real-world solutions.

Omri Azencot

Senior LecturerBen-Gurion University of the Negev

Bio:

Omri Azencot is a Senior Lecturer (Assistant Professor) in the Department of Computer Science at Ben-Gurion University of the Negev, Israel. His research focuses on deep learning, sequence models, and dynamical systems. He completed his postdoctoral research at the Department of Mathematics at the University of California, Los Angeles, under the mentorship of Prof. Andrea Bertozzi. Omri earned his PhD in Computer Science from the Technion – Israel Institute of Technology, where he was advised by Prof. Mirela Ben-Chen. Prior to that, he obtained dual BSc degrees in Computer Science and Mathematics from the Technion. His research has been supported by an ISF grant, the Adams Fellowship from the Israel Academy of Sciences and Humanities, a Zuckerman Postdoctoral Fellowship, and a Marie Sklodowska-Curie Actions International Fellowship.

Title:

Breaking Barriers in Time Series Generation: From Koopman Dynamics to Image-Based Models

Abstract:

Generating realistic time series data is crucial for various scientific and engineering applications. While generative adversarial networks (GANs) dominate this field, they suffer from instability and mode collapse. Variational autoencoders (VAEs), though more stable, remain underexplored. In this talk, I will introduce Koopman VAE (KoVAE), a novel VAE-based framework inspired by Koopman theory, which enables robust generation of both regular and irregular time series. Additionally, I will present a new paradigm that transforms time series into images, allowing the use of powerful diffusion vision models. Our approaches achieve state-of-the-art results across multiple benchmarks, bridging short- and long-range sequence modeling.

Ido Sobol

MSc StudentTechnion - Israel Institute of Technology

Bio:

Ido Sobol is an MSc student in Computer Science at Technion, supervised by Prof. Or Litany. His research focuses on 3D computer vision and generative AI, with a particular emphasis on exploring the internal mechanisms of diffusion models.

Title:

Diffusion Models in the 3D Space: From Robust Inference to New Training Procedures

Abstract:

Diffusion models have dominated generative applications. However, they still face challenges. This talk explores two key challenges:

1.Generation Artifacts: Image-based diffusion models, used for 3D tasks as Novel View Synthesis, often produce artifacts. To address these artifacts, we introduce “Zero-to-Hero”, a training-free attention filtering mechanism, enhancing quality and condition-enforcement.

Training Challenges: Training diffusion models requires large-scale datasets of the target modality, but 3D data is scarce. In our work “A Lesson in Splats”, we propose a novel training strategy that decouples the denoised modality from the supervision modality, enabling to train 3D diffusion models using only 2D supervision.

Guy Singer

Algorithmic Research LeadVisual Layer

Bio:

Guy Singer leads the research team at Visual Layer, where he develops novel algorithms for exploring, processing, cleaning, and enriching massive visual datasets. He also researches the theory of learning in neural networks at Tel Aviv University. Prior to this, he worked at the KLA Corporation, the University of Colorado Boulder, and the University of California Santa Barbara.

Title:

Garbage In, Garbage Out: How Label Noise Affects Vision Models, and What To Do About It

Abstract:

Label noise in computer vision datasets can have detrimental effects on model training, with larger datasets being needed to achieve the same post-training accuracy. We present evidence of the high prevalence of mislabels in standard academic datasets, a review of strategies for detecting and coping with these mislabels, and a release of open-source versions of these datasets that have been cleaned of their mislabels. As part of this effort, we also present LabelRank: a novel and state-of-the-art method for detecting mislabels in a distribution-agnostic manner.

Ido Marcus

Director of Research Trigo

Bio:

Ido, a member of Trigo’s founding team, manages the company’s interdisciplinary research group. Trigo specializes in computer-vision based solutions for physical retail stores, such as category leading autonomous checkout and theft detection systems. Ido’s group develops cutting edge AI in-the-wild algorithms, serving millions of shoppers around the world, at scale and in real-time.

Ido’s background as a researcher and a research engineering manager spans various fields such as deep learning, multiple-view geometry, and tracking algorithms. He holds BSc and MSc degrees in physics from the Hebrew University of Jerusalem and the Weizmann Institute of Science.

Title:

Anonymized tracking of hundreds of people across thousands of cameras, online, in autonomous retail stores

Abstract:

At Trigo we developed the world’s most advanced fully autonomous store, serving shoppers with accurate receipts in real-time. The underlying problem, namely to track all people and “understand” their interactions with products, is complex, multi-tiered and multi-faceted. Hence, solving it requires harnessing the latest advancements and exploring new frontiers across multiple algorithmic disciplines.

In my talk I’ll touch upon the various components at play, and focus on our unique multi-view multi-person tracking system, which sets a new state of the art in terms of scale and accuracy, is robust to changing environments, and adheres to strict privacy regulation.

Roie Cohen

Senior Algorithm Developer Align Technology

Bio:

Roie Cohen is a Senior Algorithm Developer at Align Technology.

His work focuses on identifying optimal NIR images for dental caries detection by combining optical-geometrical insights with traditional and AI-based computer vision techniques.
Roie earned his PhD in physics from Tel Aviv University, where he was honored with multiple awards for his biophysical research in cellular mechanics.

His groundbreaking work has significantly advanced treatments for hearing deficiencies.

Align Technology, an S&P 500 company, leads the global dental industry with its Invisalign clear aligners and iTero 3D intraoral scanners, revolutionizing dental treatment through advanced 3D imaging and CAD/CAM technologies.

Title:

Combining NIR imaging with cutting-edge AI for non-radiation diagnostics

Abstract:

Screening for dental caries is a common practice aimed at early detection and prevention of invasive treatments. Using non-ionizing-radiation methods like Near-infrared (NIR) imaging eliminates the exposure to harmful X-rays while allowing accurate diagnosis.

NIR imaging leverages the teeth's partial transparency to visualize carious lesions. However, it faces challenges like internal and specular reflections that deteriorates image quality.Our innovative solution combines classical methods with machine learning, using geometrical optics and image data.

This approach produces high-contrast, clear NIR images, enabling quick and precise diagnoses.

This allows for frequent, radiation-free patient monitoring with minimal inconvenience, revolutionizing dental diagnostics.

Andrey Gurevich

AI Algorithms Principal Engineer Mobileye

Bio:

Andrey Gurevich is an AI Algorithms Principal Engineer at Mobileye, developing computer vision systems for autonomous driving. He leads research on LLM-based object detection, unsupervised knowledge distillation, efficient fine-tuning and efficient real-time DNN architectures, driving high-impact AI projects. Andrey holds an MSc in Electrical and Computer Engineering from Ben-Gurion University, where he researched sequential anomaly detection under Prof. Kobi Cohen.

Title:

Computationally Efficient Transformer for Autonomous Driving

Abstract:

Autonomous driving requires efficient processing of multi-modal, high-dimensional data. Transformer models, while powerful, face computational bottlenecks, particularly in attention mechanisms and softmax operations. This work introduces two novel architectures: Linear Perceiver, which integrates No-Softmax Linear Attention with dimensionality reduction, and STAT (Sparse Typed Attention), which hierarchically processes multi-camera inputs using learnable latent tokens. These approaches significantly reduce complexity, achieving up to 97% fewer softmax operations and up to 20× speedup over traditional transformers, with minimal performance trade-offs. The proposed methods are scalable across hardware platforms and extendable beyond automotive applications.

Roni Goldshmidt

Senior AI ResearcherNexar

Bio:

Roni Goldshmidt is a Senior AI Researcher at Nexar, specializing in vision-language models (VLMs) and diffusion-based video generation. His work focuses on developing interpretable and efficient AI models, leveraging large-scale datasets to improve predictive accuracy and real-time decision-making. He conducts research in interpretability and multimodal learning, aiming to enhance transparency and usability in AI-driven perception systems. His work bridges academic research and practical AI applications, contributing to advancements in multimodal learning and AI-driven perception.

Title:

PixelSHAP: Extending TokenSHAP for Vision-Language Models (VLMs)

Abstract:

Building upon the TokenSHAP framework for interpreting Large Language Models (LLMs), we extend this methodology to Vision-Language Models (VLMs). Our approach, PixelSHAP, applies Monte Carlo Shapley value estimation to image data, treating different regions as token groups. This enables interpretable heatmaps that highlight the most influential visual elements in a model’s decision-making process. Evaluations on image captioning, visual question answering, and zero-shot classification tasks demonstrate PixelSHAP’s effectiveness in improving VLM interpretability. By offering structured insights into model attention mechanisms, PixelSHAP facilitates debugging, bias analysis, and enhanced trust in multimodal AI systems.

Daniel Winter

Research EngineerGoogle & The Hebrew University of Jerusalem

Bio:

Daniel is a Research Engineer at Google, where he focuses on advancing generative image models and developing innovative editing capabilities. He holds an M.Sc. in Computer Science, supervised by Prof. Yedid Hoshen, and a B.Sc. in Mathematics and Computer Science, both from The Hebrew University of Jerusalem.

Title:

Supervised Image Editing with Diffusion Models

Abstract:

Diffusion models have revolutionized image editing but often struggle to follow user intent or produce photorealistic edits. We argue that the next performance leap will come from high-quality, carefully curated supervision designed to simulate real-world scenarios. First, we introduce ObjectDrop, an object removal model trained on a small counterfactual dataset of image pairs captured before and after removing an object. Second, we present ObjectMate, a subject-driven generation and object insertion method achieving state-of-the-art photorealism with exceptional identity preservation. To create paired data, we leverage the observation that, in large unlabeled datasets, many objects reappear in diverse scenes and poses. These works showcase the power of supervised image editing.

Ofir Hadar

AI team leaderDeePathology

Bio:

Ofir Hadar is an AI team leader and applied researcher with a strong background in computer science, holding a master's degree in the field. With expertise in computer vision, deep learning, and machine learning, he brings a wealth of knowledge to his role. Ofir has vast experience in AI for computational pathology, working on many different problems from diagnostics to personalized health care and pharmaceuticals.

Title:

Segmentation by Factorization: Unsupervised Semantic Segmentation for Pathology by Factorizing Foundation Model Features

Abstract:

We introduce Segmentation by Factorization (F-SEG), an unsupervised segmentation method for pathology that generates segmentation masks from pre-trained deep learning models. F-SEG allows
the use of pre-trained deep neural networks, including recently developed pathology foundation models, for semantic segmentation. It achieves this without requiring additional training or fine-tuning, by
factorizing the spatial features extracted by the models into segmentation masks and their associated concept features. We create generic tissue phenotypes for H&E images by training clustering models
for multiple numbers of clusters on features extracted from several deep learning models on TCGA [1], and then show how the clusters can be used for factorizing corresponding segmentation masks using off-the-shelf deep learning models. Our results show that F-SEG provides robust unsupervised segmentation capabilities for H&E pathology images, and that the segmentation quality is greatly improved by utilizing pathology foundation models. We discuss and propose methods for evaluating the performance of unsupervised segmentation in pathology.

Leah Bar

Senior ResearcherAutobrains & Tel-Aviv University

Bio:

Leah Bar holds a B.Sc. in Physics, an M.Sc. in Biomedical Engineering, and a Ph.D. in Electrical Engineering from Tel Aviv University. She completed her postdoctoral fellowship in the Department of Electrical Engineering at the University of Minnesota. Currently, she is a senior researcher at Autobrains, and also a researcher in the Applied Mathematics Department at Tel Aviv University. Her research interests include machine and deep learning, image processing, computer vision, and inverse problem.

Title:

Active Learning via Classifier Impact and Greedy Selection for Interactive Image Retrieval

Abstract:

Active Learning (AL) is a user-interactive approach aimed at reducing annotation costs by selecting the most crucial examples to label. Although AL has been extensively studied for image classification tasks, the specific scenario of interactive image retrieval has received relatively little attention. This scenario presents unique characteristics, including an open-set and class-imbalanced binary classification, starting with very few labeled samples. We introduce a novel batch-mode Active Learning framework named GAL (Greedy Active Learning) that better copes with this application. It incorporates new acquisition functions for sample selection that measure the impact of each unlabeled sample on the classifier. We further embed this strategy in a greedy selection approach, better exploiting the samples within each batch. We evaluate our framework with both linear (SVM) and non-linear MLP/Gaussian Process classifiers. For the Gaussian Process case, we show a theoretical guarantee on the greedy approximation. Finally, we assess our performance for the interactive content-based image retrieval task on several benchmarks and demonstrate its superiority over existing approaches and common baselines.

Lev Ayzenberg

PhD candidateTel Aviv University

Bio:

Lev Ayzenberg is a Direct Track PhD candidate at Tel Aviv University's TAD graduate program, where he researches deep learning and computer vision under the supervision of Prof. Raja Giryes and Prof. Hayit Greenspan. His research focuses on foundational models and few-shot learning. Currently, he works as a Deep Learning Algorithm Engineer at RADLOGICS, developing medical imaging solutions. Prior to his academic career, Lev served in the IDF's Matzov Unit as a hardware engineer and kernel security researcher.

Title:

ProtoSAM - One shot medical image segmentation with foundational models

Abstract:

We introduce ProtoSAM, a fully automated framework for one-shot medical image segmentation that combines Prototypical networks with the Segment Anything Model (SAM). The method uses ALPnet prototypical network with a DINOv2 encoder to create an initial coarse segmentation mask, from which prompts are extracted and fed into SAM. Extensive validation on CT, MRI, and polyp image datasets demonstrates state-of-the-art results in multiple scenarios. Notably, untrained (zero-shot) ProtoSAM matches or exceeds the performance of existing trained methods, with potential for further improvements through self-supervised single-example fine-tuning of the encoder.

Alexandra Dana

Computer vision lead engineer Samsung

Bio:

Alexandra is a seasoned computer vision algorithm engineer with more than 20 years of experience in the industry.

In the last ten years she has been leading various computer vision projects at Samsung for its automotive and mobile SoCs.

She holds a PhD in Bioinformatics from Tel-Aviv University, MSc in Electrical Engineering from Tel-Aviv University and a BSc in Computer Engineering from the Technion and the author of numerous academic papers and patents.

Title:

Do More With What You Have: Transferring Depth-scale from Labeled to Unlabeled Domains

Abstract:

Adjusting pre-trained absolute depth predictors to new domains is a task with significant real-world applications.

This task is specifically challenging when images from the new domain are collected without ground-truth depth measurements, and possibly with sensors of different intrinsics.

We suggest an novel solution for this challenge and successfully demonstrate it on the KITTI, DDAD and nuScenes datasets, while using other existing real or synthetic source datasets, achieving comparable or better accuracy than other existing methods that do not have access to target ground-truth depths.

Hila Levi

Research engineerGeneral Motors

Bio:

Hila Levi is a Senior Researcher in the R&D group at General Motors. Hila holds an M.Sc. from Tel-Aviv University and a Ph.D. from the Weizmann Institute, specializing in computer vision. Previously, Hila held algorithm development positions at Vimeo, Elbit and Applied Materials, gaining extensive experience in both academic and industrial settings. Hila’s areas of interest include computer vision, language models, and their interactions, and her current research focuses on advancing machine perception and scene understanding in the automotive industry.

Title:

Open-Vocabulary Object-Based Image Retrieval

Abstract:

Open-vocabulary object-centric image retrieval focuses on retrieving images containing objects of interest based on an open-set text query. This task has applications in targeted performance analysis, hard example mining, and more. In our works, “Object-Centric Open-Vocabulary Image Retrieval with Aggregated Features” (BMVC 2023) and “FOR: Finetuning for Object-Centric Open-Vocabulary Retrieval” (WACV 2025), we address the tradeoff between accuracy and efficiency in object-centric retrieval. We propose effective methods for generating compact yet expressive image representations and introduce fine-tuning strategies that leverage a limited set of labeled categories to enhance retrieval accuracy across all categories, particularly useful in low-data regimes.

Eva Seknazi

Tech lead algorithm developerApplied Materials

Bio:

Eva Seknazi is an algorithm tech lead at Applied Materials in the Process Diagnostics and Control business unit. In this role, Eva is responsible for leading the development of cutting-edge metrology algorithms, both in classical and deep learning domains. She has obtained a PhD from the department of Materials Science and Engineering at the Technion – Israel Institute of Technology.

Title:

Beyond Nanoscale: Novel SEM Image Segmentation Technique for Accurate Metrology

Abstract:

The chip manufacturing process heavily relies on accurate segmentation of semiconductor patterns to ensure the quality and efficiency of advanced chips. This talk addresses the challenges involved in segmenting nano-scale features in scanning electron microscope images of semiconductor patterns. We propose a foundation model, trained on a large dataset, and specifically designed to tackle the unique challenges present in this domain. We demonstrate the high performance and robustness of our algorithm, establishing it as a groundbreaking tool for the semiconductor industry. With its versatility and ease-of-use, our approach paves the way for advancements of semiconductor metrology and chip manufacturing.

Shahaf Finder

Ph.D. CandidateBen-Gurion University of the Negev

Bio:

Shahaf E. Finder is a co-founder of an AI startup in the manufacturing space and a PhD candidate at Ben-Gurion University, specializing in machine learning and computer vision. In the final stages of his doctoral work, he focuses on designing and optimizing efficient neural network architectures.
Abstract: In recent years, there have been attempts to increase the kernel size of Convolutional Neural Nets (CNNs) to mimic the global receptive field of Vision Transformers’ (ViTs) self-attention blocks. That approach, however, quickly hit an upper bound and saturated way before achieving a global receptive field. In this talk, I will present our recent work that leverages the Wavelet Transform (WT) to obtain very large (and even global) receptive fields in CNNs. We will also discuss how the WT helps CNNs respond better to low frequencies in the input, and the benefit of it with regard to robustness and response to shapes.

Title:

Wavelet Convolutions for Large Receptive Fields

Abstract:

In recent years, there have been attempts to increase the kernel size of Convolutional Neural Nets (CNNs) to mimic the global receptive field of Vision Transformers’ (ViTs) self-attention blocks. That approach, however, quickly hit an upper bound and saturated way before achieving a global receptive field. In this talk, I will present our recent work that leverages the Wavelet Transform (WT) to obtain very large (and even global) receptive fields in CNNs. We will also discuss how the WT helps CNNs respond better to low frequencies in the input, and the benefit of it with regard to robustness and response to shapes.

Assaf Arbelle

Senior Research Scientist and ManagerIBM

Bio:

Assaf Arbelle (PhD) is a Senior Research Scientist and Manager of the Multimodal-AI group at IBM Research.

Assaf leads a team of 13 scientists with expertise in the fields of Computer Vision, Speech, NLP, and more. The team focuses on Multimodal Foundation Models, working to develop the next generation of AI systems capable of handling multiple tasks across various modalities. Assaf’s research interests include the intersection of Computer Vision and Language Models, as well as broader topics in Machine Learning, Speech and Audio, and general AI.

The work at IBM Research emphasizes advancing AI for business, implementing innovative methods and tools to accelerate real-world use cases.

Title:

Training a Large Vision-Language Model - Insights and lessons learned from building IBM’s Granite-Vision VLM

Abstract:

With the exponential growth of Large Language Models came a rise in the joint modality paradigm of Large Multimodal Models (LMMs) which expand the LLM’s reasoning and understanding capabilities to more modalities such as Images and Audio. This talk will present the work done at IBM Research to build Granite Vision, a lightweight large language model with vision capabilities, specifically designed to excel in enterprise use cases, particularly in visual document understanding. We will discuss the network’s architecture, the data used to train the model, and the technical details of training the model. We will also present compelling results surpassing SoTA models on enterprise related tasks.

Sivan Doveh

PhD student Weizmann Institute of Science

Bio:

Sivan Doveh is currently a Ph.D. Candidate at the Weizmann Institute of Science. Her research focuses on enhancing the core abilities of multimodal models, such as compositional reasoning and in-context learning. She publishes in conferences such as ECCV, ICCV, CVPR, and NeurIPS, where she also serves as a reviewer. Sivan holds an M.Sc. in Electrical Engineering from Tel Aviv University and, in parallel to her Ph.D. studies, previously worked as a research scientist at IBM Research.

Title:

Teaching VLMs to Localize Specific Objects from In-context Examples

Abstract:

Vision-Language Models (VLMs) excel in various visual tasks but struggle with context-aware object localization, a crucial ability in ambiguous scenarios. We address this gap with a few-shot personalized localization task, where a model learns to localize objects from a small set of annotated examples. To enhance contextual reasoning, we fine-tune VLMs using curated video object tracking data, simulating instruction-tuning dialogues. Additionally, we introduce a novel regularization technique that replaces object labels with pseudo-names, enforcing reliance on visual context. Our approach significantly improves few-shot localization while preserving generalization, establishing the first benchmark for personalized localization and advancing context-driven vision-language modeling.

Ido Yerushalmy

Science Team LeadAmazon Prime

Bio:

Ido Yerushalmy leads a Science team at Amazon Prime Video Sports in Tel-Aviv, focusing on developing deep learning solutions for sports broadcast analysis. His team pioneers advanced algorithms for game understanding and predictive analytics in live sports content. Previously, Ido specialized in computer vision applications, particularly pose estimation and object detection. Prior to that he worked at CorePhotonics and Samsung, where he developed 3D stereoscopic vision systems and medical image analysis solutions. Ido holds an MSc in Computer Science from the University of Haifa and brings over a decade of experience in applied machine learning and computer vision technologies.

Title:

Amazon prime video

Abstract:

This lecture will present the advanced analytics and machine learning-powered features developed by Amazon to enhance the live viewing experience for Thursday Night Football on Prime Video. Leveraging player tracking data, the team developed three novel features: Defensive Vulnerability, which identifies defensive weak spots before the snap; Pressure Alerts, a deep learning model that predicts quarterback pressure; and Coverage Prediction, which forecasts man vs. zone coverage. These features harness Amazon's cloud and edge computing capabilities to deliver real-time insights to both casual and avid fans. The results highlight how sports broadcasting is shifting towards a data-driven approach powered by the latest advancements in artificial intelligence.