Please note that this agenda and scheduling is subject to changes

Diffusion models are considered to be the state-of-the-art tool today for image generation guided by text. In this work we demonstrate how this tool can be used for 3D shape generation and for creating high-res textures for existing shapes. In addition, we show how diffusion models can be used for image reconstruction and editing .

Recent advances in visual generation are almost unimaginable. This leap forward in capabilities has happened thanks to several factors, including the adoption of the diffusion approach, elaborate data collection, and more. I argue that one of the main forces to drive visual generation advancements is Language: recent Large Language models encapsulate astounding amounts of knowledge, since data collection is simpler for text compared to images or video. In this talk, I scan five of my recent publications, spanning generative capabilities from facial and sketch images, to general images and 3D human motion. Through these publications, I portray how Large Language Models (and especially CLIP) become more and more significant for controlling and improving visual generation.

Wouldn't we prefer segmentation over detection? a pixel level detection is preferable. scale sensitive, no anchors, no NMS. Let's see how this can be achieved.

Ofir will talk about how Lightricks implemented generative AI and the latest developments in "image transformation."

Computer vision now heavily relies on "foundation models", text-and-image representations trained with massive data in a task-agnostic way, and then used for downstream tasks. But how can we make them useful for our own tasks? You may want to search for a photo of your own pet in your photo album, or generate an image of your friend in a funny situation.   I will discuss our work on personalizing text-vision models, by teaching models new words and concepts. I'll show applications to personalized ranking, segmentation and image generation.  

Explainability is an ever-growing field of research, aiming to demystify the black box that is deep learning. In recent years, the naturally interpretable attention mechanism has become one of the most common building blocks of neural networks, allowing us to produce explanations intuitively and easily.
However, the applications of such explanations beyond the scope of accountability and interpretability remain limited.
In this talk, I will present my latest research on leveraging attention to significantly improve the accuracy and robustness of state-of-the-art large neural networks with limited resources. This is achieved by directly manipulating the attention maps based on intuitive objectives and can be applied to a variety of tasks ranging from object classification to image generation.
 

This presentation delves into the field of speech analysis, focusing on the topic of attacks on voice conversion models. I will discuss the current state of the art in voice conversion technology and the various challenges faced by these models. I will also provide insight on how to create effective online voice conversion model.

The recent advancement in deep learning has re-opened the field of acoustics sensing for unique defense related
tasks such as drone detection, gunshot detection, ultrasonic sensing and more. Learn about the recent advances in this field and how the IAI team uses deep learning to overcome the past limitations of sound sensing to build innovative new products.

The success of the best products in the defense sector is usually based on well-defined system requirements, even if highly complicated. "Well defined requirements" refers especially to the borders and interfaces of these systems with the customer and/or the environment.
On this talk, however, we focus on cases where these requirements and borders could not be clearly set up front, and  perhaps even should not. Such cases arise when a solution is requested to be integrated deep into an existing operational ecosystem in which the customer already operates. This is even more acute when we are dealing with AI-based solutions, which are notoriously hard to maintain after deployment. In these cases, the best option is to take the agile development process to the extreme, and agilely evolve the system requirements together with the actual system.
Three such success cases from Rafael's joint AI-hubs will be presented, in which a successful AI-based solution was developed together with the customer and deployed into his operational environment in a short time. The quick  feedback received from the customer was imperative to adjust the solution in response to changes in the operational scenario.

The intelligence directorate in the IDF, and in it the 8200 unit, has a vital mission to implement digital transformation and utilize artificial intelligence. In this age of overwhelming information from various sources, a new approach is required to discover the most insightful and trustworthy data and breaking human barriers in the intelligence process.

Personalization in gaming is the concept of tailoring the best gaming experience to each player based on a variety of parameters such as their preferences, habits, level, etc..  Personalization goal is to increase player engagement and create a more immersive and satisfying gaming experience. In this presentation, we will discuss the importance and benefits of personalization in gaming, and how Playtika is leveraging a human-machine team to deliver personalized experiences to its players through scalable and generic digital products. We will cover topics such as data standardization, user modeling, and human-machine collaboration in the context of Playtika's digital platforms. This presentation is intended for professionals and technical audiences with an interest in gaming, personalization, and human-machine collaboration.

Toxic content such as misinformation, cyberbullying, harassment, and child sexual abuse material has become an integral part of the online ecosystem. To ensure compliance with community standards, millions of pieces of potentially harmful content are flagged daily by social media platforms using a combination of automated and human reviews. Despite this, content moderation remains a pernicious problem for online platforms. 
 
Most platforms today use centralized AI-enabled models, whereby a common algorithm scans all posted content.  Given the wide variety of groups and communities that participate in online interactions, each with its own parlance, behavioral norms, and topics, a common algorithm applied indiscriminately across all users ignores the context and leads to errors.  
 
In this talk we propose an alternative, decentralized framework for AI-enabled content moderation that uses a more tailored algorithm augmented with a contextual layer adapted to the type of content that a given community finds acceptable.  We show that such an approach can lead to fewer false positives and decrease the load on human moderators. We conclude with thoughts on how this approach can enable self-governance by online communities and help address the issues of censorship and control raised against technology platforms and government regulators.  
 

The energy sector is facing an ever-increasing range of challenges coming with decarbonization and decentralization. In many use cases, AI solutions could bring an efficient answer.

Machine learning can extend portfolio of support functions used by system operators, power system engineers, or data analysts. ČEPS (the Czech electricity transmission system operator) successfully piloted three use cases and plans to adopt machine learning operations in the following years. Power systems appears to be a suitable domain for extensive use of artificial intelligence methods. ČEPS is also active in the European sector organization of TSOs (ENTSO-E), which seeks to find common high-value use cases for applying AI to build an enhanced control room of the future.
 

Artificial intelligence (AI) has great potential to transform many aspects of health care – but despite this promise has only been implemented in a limited number of settings and in a limited way. This talk begins by discussing the potential of AI to transform health care in areas such as disease diagnosis and treatment, disease monitoring, patient monitoring and clinical outcome prediction, health care operations, and health care administration. We then outline the many reasons why this potential has not been realized. As a microcosm illustrating the state of AI applications, we describe a number of projects carried out at Lucile Packard Hospital Stanford. We conclude with thoughts about what is needed in order for AI to be successfully integrated into health care systems and the next steps that can be taken to advance implementation.

Our bodies are built out of billions of cells, each containing DNA and RNA with information on our health and disease states. AI has increased in popularity as a useful approach to solving many complex bio-computational challenges. We show that we can employ advanced bioinformatics in order to identify early changes that can lead to diseases. Our work will lead to the seamless integration of bio-digital breakthroughs in medicine.

Migraine is one of the most common neurological disorders, affecting over 1 billion people worldwide [1], ranked second among the causes of disability globally [2] and there are no clinical biomarkers for migraine attacks [3]. While the term migraine is usually associated with the few hours of the pain phase, the physical and clinical changes associated with the migraine phenomena may start a few days prior to the pain phase and may last a few days afterward, therefore, can be divided into 3 stages [3]; (1) The migraine prodrome: Premonitory symptoms that occur before the headache phase, (2) The headache phase: lasts 4-72 hours, (3) The migraine postdrome: Residual symptoms after the pain phase that may last up to 72 hours.Understanding migraine attack phases can help anticipate and manage the symptoms. Therefore, our aim is to demonstrate the ability to distinguish between the different migraine stages using personal physiological metrics based on continuously monitoring by smartwatches. Using data collected continuously from smart watches and cellular phones, we can identify precursors and conditions that are correlated with different stages of migraine attacks.
 

Contributors: Moria Dagan, Adar Frenkel, Aviv Balassiano, Tamar Spector, Gad Resef, and Eddie Aronovich 

The healthcare industry is facing new and increasing challenges in terms of patient care, with increasing patient populations and limited resources. Artificial intelligence (AI) has the potential to revolutionize healthcare, but it requires unique considerations to benefit patients in practical and meaningful ways. This talk will explore how keeping the patient at the center of AI development can ensure ethical and effective use of the technology and how patients can benefit from AI, focusing on the wound care domain. With a growing demand for efficient and effective wound care, AI-based wound monitoring solutions have emerged as a promising solution for patients. These solutions use advanced algorithms and sensors to accurately and quickly assess wounds and increase patient comfort and convenience. The practical application and impact of this solution on patients will be demonstrated through a real-world study we conducted in collaboration with Johns Hopkins Hospital, using our Minuteful for Wound product on a vulnerable population. This example will provide you with insights into the potential of AI in remote patient monitoring and the importance of patient-centered approaches in its development and implementation.

Despite extensive technological advances in recent years, objective and continuous assessment of physiologic measures after vaccination is rarely performed. To evaluate the safety profile of COVID-19 vaccines, we analyzed their short-term effects by using data from, first, a retrospective cohort of 250,000 random members of the second-largest health-care organization in Israel and, second, a prospective cohort (the PerMed study) of 4698 participants that wore smartwatches and filled questionnaires. Our findings further support the safety of the vaccines from subjective and objective perspectives and underscore the need for integrating wearables in clinical trials due to their higher sensitivity.

We are facing a climate emergency.  Temperatures are rising, rainfall patterns are changing, and sea level is rising.  As a result, we are seeing a dramatic increase in natural disasters related to both dry weather (heat waves, droughts, wildfires) and wet weather (river floods, storms, coastal flooding), with the Global South suffering most from climate change, even though the Global North is responsible for the crisis. How can AI be used for both adaptation and mitigation of climate change in the near and far future?  AI brings tools that can be used to distill raw data into actionable information, to improve predictions in sectors such as energy, agriculture and extreme weather.  AI can optimize complex systems such as heating and cooling systems or transportation schedules.  AI can be used for accelerating scientific modeling and discovery using machine learning with physics-based constraints.  However, at the same time we need to ensure that AI will not have negative impacts on the Earth’s climate.  Examples of where AI can promote climate action will be presented in the talk.

If you've considered using AI to do good, you may have run into the challenge of figuring out where you could help the most. This talk will describe some of the methodology used by research teams at Google when facing the challenge of choosing which problems to work on, and describe some examples of existing projects such as flood forecasting, traffic light optimization, post-disaster damage assessment and more. In addition, the talk will explore more in-depth the ways research teams at Google are using AI to help detect and predict wildfires

Soil provides ecosystem services, supports human health and habitation, stores carbon and regulates emissions of greenhouse gases. Unprecedented pressures on soil from degradation and urbanization are threatening the agro-ecological balance and food security. It is important that we learn more about soil to sustainably manage and preserve it for future generations. Soil is a complex system that is extremely variable in physical and chemical composition in space and time. It is formed from exposed masses of partially weathered rocks and minerals of the earth's crust. Soil spectroscopy across the VIS-NIR-SWIR-TIR (0.4-14m) spectral region is less complex than the soil itself and hence can reduce the uncertain dimension of the soil system. Furthermore, a proxy data-mining approach for extracting spectral information that refers to soil attributes is possible and has gained much attention over the past decade in the precision agriculture discipline. A suitable data-mining AI based algorithm enables extracting a mixture of spectral information that is hidden in the spectrum: it cannot be seen or extracted with the naked eye or even by ordinary computation. As the spectral data-mining approach is statistically driven, a large set of samples ("big data") is required to establish a precise soil spectral model for a given soil attribute. Accordingly, over the last 5 years, extensive effort has been devoted in building, sharing and using soil spectral databases (also termed soil spectral libraries—SSLs) on national, continental and global scales. SSLs consist of metadata (location, geology, climate, topography, horizon, area description and location) plus chemical and physical attributes from wet laboratory analyses. This together with the soil spectral information the SSL is used to derive soil attributes from a complex system without having to bring it to the laboratory; as such, it is important for the forthcoming super spectral and hyperspectral sensors. As the soil spectral user community expands, the demand for large SSLs from many geographical areas on every scale intensifies. Thus, several regional SSLs have been evaluated (e.g., Africa, Brazil, USA and Europe) while more recently, a worldwide effort has been devoted to creating a global SSL. Recently a new era in space where hyperspectral sensors became operational, exploiting the SSLs  for soil quantitative mapping takes an interesting direction. The current lecture will provide an update overview of  soil spectral analyses to foster soil monitoring under the precision agriculture era .

Data Science, and in particular Data-driven research (DDR) is proliferating in light of the development of artificial intelligence (AI) technologies, which enables to process enormous amounts of data, redefining research capabilities. Governments and industry are the primary sources of big, structured data, which they commonly generate in the course of their routine activities. Academic institutions, whose research activities were, in the past, largely autonomous given sufficient funding, have become increasingly dependent on such third-party data to fuel their DDR. This development stands to transform the interrelations between the academy-industry- government (AIG) sectors and has far-reaching implications for policies aimed at facilitating scientific research and innovation. In this talk we will review the emerging challenges with respect to the technology transfer between academia and industry when external, 3rd party data is involved, and will point on the potential role of governments in facilitating it for the benefit of society as a whole.

In this talk, we discuss several gaping problems of modern LLMs that became particularly salient after the release of ChatGPT. 
How can we make models like ChatGPT truly multilingual? We discuss a set of performance evaluation experiments and the discrepancies between the distribution of languages in the training data and the distribution of languages in terms of their native speakers across the globe. We illustrate how crowdsourcing can amend these challenges using the example of WMT’22 Shared Task on Large-Scale Machine Translation Evaluation for African Languages. 
Another question we have to answer is how LLMs can work in narrow hands-on applications reliably and ethically. With a series of experiments with LLMs and multimodal generative models, we try to illustrate how modern crowdsourcing can facilitate prompt engineering and even assist in fine-tuning a model for a more narrow use case. Finally, we discuss the demand for more skilled and committed data labelers that arises as LLMs get broader adoption.
 

Abstraction is a core tenet of human cognition and communication. Yet, interpreting and grounding abstraction expressed in natural language has not yet been systematically studied in NLP. In this talk, I present our recent work on setting the foundation for a systematic study of processing and grounding abstraction in natural language, and in particular, computational abstraction that hinges on human perception and cognition. First, I draw an explicit link between human language communication and computational thinking skills. Next I present a novel elicitation method we devised via Hexagons, a 2D instruction-following game, and a crowdsourced benchmark we collected based on it, with over 4k naturally-occurring visually-grounded instructions, rich and diverse in their levels of abstraction. Our experiments with these data show that contemporary models and modelling practices are substantially inferior to human performance, and that, crucially, models' performance is inversely correlated with the level of abstraction, with less satisfying performance at the higher abstraction levels. These findings are consistent across models and setups, confirming that NL phenomena which involve abstraction present a challenging domain of investigation, deserving further attention in NLP/AI research in the quest for more human-like human-computer interaction.
 

Dual encoders are now the dominant architecture for dense retrieval. Yet, we have little understanding of how they represent text, and why this leads to good performance. Here we shed light on this question via distributions over the vocabulary. We propose to interpret the vector representations produced by dual encoders by projecting them into the model's vocabulary space. We show that the resulting distributions over vocabulary tokens are intuitive and contain rich semantic information. We find that this view can explain some of the failure cases of dense retrievers. For example, the inability of models to handle tail entities can be explained via a tendency of the token distributions to forget some of the tokens of those entities. We leverage this insight and propose a simple way to enrich query and passage representations with lexical information at inference time, and show that this significantly improves performance compared to the original model in out-of-domain settings.
Joint work with: Ori Ram, Liat Bezalel, Adi Zicher, Yonatan Belinkov and Jonathan Berant.
 

Large Language Models (LMs) are Improving at an incredible pace, but some inherent limitations of Transformer-based LMs remain problematic, e.g., generating factually ungrounded text, processing only bounded length text sequences, and more. In recent years, a multitude of successful solutions for these issues has been proposed. However, these involve specialized architectures and cannot be applied to off-the-shelf large LMs, rendering them irrelevant to models coming off the mainstream large-LM production line. In this talk I will present new methods for tackling the above drawbacks and others, which are applicable to any off-the-shelf large LM, without further training it. By addressing inherent LM weakspots in a model agnostic approach, we hope to alleviate them in a widespread manner, in existing and future leading LMs.

Incorporating humor in systems can enhance the user experience, attention and retention, as well as user’s rapport with the product. However, despite being studied for millennia, humor is computationally not well understood, often considered an AI-complete problem. Thus, much of the effort towards automating humor was done on very narrow and well-defined tasks. 
In this talk, I will present three novel computational humor works (and what they taught us):

In the first task we introduce a novel setting in humor mining: automatically detecting funny and unusual scientific papers [1]. We are inspired by the Ig Nobel prize, a satirical annual prize celebrating funny scientific achievements (example winner: “Chickens prefer beautiful humans”). 

In the second we explore playfulness in human-conversational agents (CAs) interactions [2]. While CAs are well-suited for task-oriented interactions (“Call Jason”), playful requests are often beyond their capabilities (“What’s your favorite color?”). This causes dissatisfaction and abandonment, destroying the rapport with the agent. Based on humor theories and real traffic data from Amazon Alexa, we present a taxonomy of playful requests and discuss ways to deploy it. 

In the third, we close the loop and explore ways to automate the process of generating appropriate responses to playful shopping requests [3].

We hope that insights gained from these studies will help tackle long-standing challenges in AI and social science.

[1] Shani, C., Borenstein, N., & Shahaf, D. (2021, August). How Did This Get Funded?! Automatically Identifying Quirky Scientific Achievements. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) (pp. 14-28).

[2] Shani, C., Libov, A., Tolmach, S., Lewin-Eytan, L., Maarek, Y., & Shahaf, D. (2022, April). “Alexa, Do You Want to Build a Snowman?” Characterizing Playful Requests to Conversational Agents. In CHI Conference on Human Factors in Computing Systems Extended Abstracts (pp. 1-7).

[3] Shapira N., Kalinsky O., Libov A., Shani C. & Tolmach S. (2023). Evaluating Response Generation to Playful Shopping Requests. In European Conference on Information Retrieval. Springer, Cham.
 

Learning algorithms have been an important cornerstone in numerous modern application domains, and are a key ingredient in many technological eco-systems. Examples include A/B testing,  recommender system design, pricing analytics, trade execution algorithms, and online health monitoring, to name but a few. Many such algorithms are deployed and run at massive scale and are part and parcel of the AI revolution we are all witnessing.  The academic literature, from which many of these algorithms and ideas originate, has mostly focused on what can be referred to as “stationary” environments, where ambient operating parameters and conditions remain stable. This is rarely the case in many real-world settings. In this talk we will illustrate some design principles that allow classical online learning approaches to be adapted to changing environments.  The ideas will be illustrated in a simple stylized setting (albeit one that has been widely adopted in the tech sector) but are applicable more broadly. 

Financial fraud ML is characterized by a setting which is both adversarial and suffering from delayed feedback. Fraud is often reported and reviewed by humans, which means that time goes by until it is noticed; during this time, fraudsters are generating new types of attacks. To build models relevant for the present, we need to train on data which is not fully labeled. Standard semi-supervised learning usually partitions the dataset to a small subset of labeled data, and a larger subset of unlabeled data. In this setting, however, all data is labeled only to some extent, without a clear partition. In this talk we will review some theoretical aspects of this, as well as graph-based models for addressing this.

The watchlist feature in e-commerce allows users to track items over time and is an important aspect of the shopping journey. This talk introduces a new watchlist recommendation task, which aims to predict the next item the user will click on by ranking items in the watchlist. We propose a recommendation model, Trans2D, that is based on the Transformer architecture, and evaluate our model using a large-scale watchlist dataset from eBay. We show our model is superior to other state-of-the-art models.

Are you looking for a way to make your fraud detection models more accurate and reliable? In this lecture, we’ll explore how Generative Adversarial Networks (GANs) can help you tackle the various challenges of AI-driven fraud detection, such as imbalance, lack of positive labeled data, and concept drift. We’ll discuss how GANs can be used to generate synthetic data, and how it can be applied to fraud detection models.

Foundational work in machine learning theory has long established the relationship between and generalization and the (in)capacity to memorize randomly assigned labels. The link between the two remains even in the deep learning era as we confront models capable (given sufficient training time) of memorizing arbitrary labels. In this talk I’ll introduce several works centered on the theme of memorization and label noise, including (i) a work that suggests that if at some point along a training run we have high accuracy on clean training data but haven’t memorized the (incorrect) noisy labels, then it must be the case that we have already generalized to the underlying population. I’ll go on to discuss some empirical work aimed at empirically characterizing the phenomenon of memorization deep learning, including efforts to identify mislabeled points and to identify the neural network structures implicated in memorization.

We present a method to overcome a common challenge in deep learning: efficiently handling dynamically changing inputs. Consider for instance an AI writing assistant making suggestions as the document is edited, or an agent evaluating which of the next moves will lead to the most favorable board game position. Conventional neural networks must fully process each input, even if they differ by a small amount. We propose a novel deep learning architecture that supports incremental computation, allowing it to continuously update predictions as inputs change while only partially recalculating its computational graph. We demonstrate our method through its application to one of the most challenging problems in quantum physics: finding the ground state of quantum many-body systems. The critical step of neural network-based approaches to such problems is calculating physical quantities known as “local energies”, which involves repeated evaluations of the spin particles under small changes. By leveraging our incrementally computable neural network, we observe a linear acceleration (growing with the number of simulated particles), achieving >10X reduction in computational resources.

Graph neural networks (GNNs) are widely used for modeling complex interactions between entities represented as vertices of a graph.  Despite recent efforts to theoretically analyze the expressive power of GNNs, a formal characterization of their ability to model interactions is lacking.  My talk will present a recent work aiming to address this gap.  Formalizing strength of interactions through an established measure known as separation rank, I will quantify the ability of certain GNNs to model interaction between sides of a given partition of vertices.  The result will reveal that the ability to model interaction is primarily determined by the partition’s walk index — a graph-theoretical characteristic defined by the number of walks originating from the boundary of the partition.  As a practical application of the theory, I will present an edge sparsification algorithm named Walk Index Sparsification (WIS), which preserves the ability of a GNN to model interactions when edges are removed.  WIS is simple, computationally efficient, and markedly outperforms alternative methods in terms of induced prediction accuracy. More broadly, it showcases the potential of improving GNNs by theoretically analyzing the interactions they model.

The presented work was in collaboration with my students Noam Razin and Tom Verbin.
 

Consider a finite sample from an unknown distribution over a countable alphabet. Unobserved events are alphabet symbols which do not appear in the sample. Estimating the probabilities of unobserved events is a basic problem in statistics and related fields, which was extensively studied in the context of point estimation. In this work we introduce a novel interval estimation scheme for unobserved events. Our proposed framework applies selective inference, as we construct confidence intervals (CIs) for the desired set of parameters. Interestingly, we show that obtained CIs are dimension-free, as they do not grow with the alphabet size. Further, we show that our CIs are (almost) tight, in the sense that they cannot be further improved without violating the prescribed coverage rate. We demonstrate the performance of our proposed scheme in synthetic and real-world experiments, showing a significant improvement over alternatives.

Drug discovery and development is a lengthy and expensive process with more than a 90% failure rate. Currently, most methods for assessing preclinical safety of a new drug candidate rely on animal testing, which fails to reliably predict the interaction of the drug with the human body. More than that, legislation in the US and EU aims to stop animals from being used for scientific purposes. Indeed, the FDA recently canceled the mandating of animal testing in the approval process of a new drug. Thus, the development of new methods for the prediction of clinical safety is a global, major unmet need. In this talk, I'll present a BIO-AI prediction platform for clinical safety. The platform integrates state-of-the-art machine learning, with patient-on-a-chip (miniaturized human organs), real-time nano-sensing, and stem-cell genomic diversity technologies. The platform performs millions of AI-guided interactions between genetically diverse patient-on-a-chip and potential drugs. While observing the interactions, the platform generates unique data, highly indicative of drug toxicities. The platform-generated data is used for dynamically training multimodal deep learning to predict drug safety in a multi-task fashion. The platform is currently used by big pharmaceutical companies for the safety assessment of new drug candidates. It will be used in the future to generate recommendations for personal safety and treatment using approved drugs.

The full potential of preventive medicine has yet to be unlocked. 
There are challenges in the way for physicians and health plans: it’s hard to identify the right members for valuable interventions, members are not engaging and the operation lacks precision, efficiency and effectiveness. 
As a result, members enter proactive clinical programs too late and avoidable health incidents are missed. 

Traditional solutions attempt to explain why patients are at risk. 
However, is it sufficient? For efficient preventive care one should identify which treatment is best for each patient, as well as it’s effectiveness.

In this talk, we will present a Proactive AI platform that identifies the most impactable members and provides actionable insights.
By advanced explanations and causal inference methodologies, and by utilizing multi-modal models, we provide for each patient at risk the link between specific clinical events and their impact and allow to assign each patient the right and most impactful intervention based on the AI-based actionable information.

In this talk Professor Eyal Zimlichman MD, MSC provides a review of how the future hospital must take shape through digital health tools utilized at ARC, the innovation arm of Sheba Medical Center (Sheba), ranked for four consecutive years as a top-10 medical center in the world by Newsweek. Professor Zimlichman will describe the Future of Health (FOH) community, which is a proactive forum of leaders working to influence the course of global health through creation of a shared vision. The ARC global initiative then works to turn the FOH vision into reality among the member organizations. To enable this transformation, ARC is based on four fundamental principles which include digital health, open & organic innovation, international collaborations, and a structured innovation methodology.
He continues with the variety of technologies developed and deployed at Sheba in artificial intelligence (AI), specifically in the field of digital radiology, digital pathology, AI decision support platform for clinical evidence, multi-omics, preventing medication errors, and an AI hospital wide control tower. Prof. Zimlichman concludes that all stakeholders must take a proactive approach in order to see the changes take shape earlier. These include policy makers, payers, care recipients, and caregivers who all need to be at the table to enable and deliver effective transformation.
 

In recent years, Deep Learning has made significant contributions to a variety of domains, such as image, video, and text. However, the significant field of electronic signals, which includes many fields such as communications, radar, medical devices (such as EKGs), navigation (such as IMUs) and acoustics, is often overlooked in AI conferences. This session, co-hosted with Elbit Systems ISTAR & EW, will discuss the main challenges in this field and the differences from other well-researched areas.

Detection and spectro-temporal localization of communication transmissions are crucial for tasks related to understanding, managing, and protecting the RF spectrum. Achieving this goal for wideband spectrum in real-time is a challenging problem. Existing methods are not capable of detecting multiple RF emissions in real-life scenarios, where the spectrum is heavily occupied. Here, we will present the problem, discuss the main challenges, overview solutions available in the open literature, and suggest other solutions based on our own experience.

Direction of arrival (DoA) estimation is a fundamental task in array processing. A popular family of DoA estimation algorithms are subspace methods, which operate by dividing the measurements into distinct signal and noise subspaces. Subspace methods, such as MUSIC and Root-MUSIC, require the sources to be non-coherent, and are considerably degraded when this does not hold. Recent years have witnessed a growing interest in using deep learning for DoA estimation. Here, deep neural networks learn their mapping from data and replace classic algorithms that are based on specific statistical modelling. However, blackbox neural networks do not share the interpretability, reliability, and flexibility of statistical model based algorithms, such as subspace methods. This talk will present a middle ground approach, following the emerging methodology of model-based deep learning. We will review how deep learning tools can be used to empower subspace methods, instead of replacing them.

By adopting such a hybrid model-based/data-driven approach, we show that one can realize DoA estimators which learn how to divide the observations into distinguishable subspaces. The resulting algorithms leverage data to cope with coherent sources, low SNR and limited snapshots, while preserving the interpretability and the suitability of classic subspace-based DoA estimation.

The collection of high-quality labeled medical data is often costly and time-consuming, limited by access to expensive measurement devices and domain experts, which eventually lead to small dataset sizes. Self-supervised learning (SSL) has emerged as a promising approach for training deep learning models in such scenarios, where a dataset consists of a small amount of labeled data and a large amount of unlabeled data. In this study, we present an SSL-based approach for continuous measurement of gait speed using a smartwatch. Gait speed is a crucial predictor of disease progression and overall health in individuals with neurodegenerative diseases such as Multiple Sclerosis (MS), Parkinson's Disease (PD), and Alzheimer's Disease (AD). We used millions of hours of accelerometer data collected from smartwatches to create a pre-trained SSL model for multivariate time series, which we then fine-tuned for the task of gait speed estimation based solely on accelerometer data from a wrist-worn smartwatch. Our results demonstrate the effectiveness of SSL in training deep learning models for continuous measurement of gait speed in neurodegenerative disease populations.

Speech and audio modeling is a challenging task due various reasons. First, such signals are extremely long, the common sampling rate for speech signals is 16kHz, while for music is 44.1kHz. Second, unlike text, speech and audio representations are continuous in nature, which makes modeling and sampling difficult. In this talk, I will present an alternative view for generative speech and audio modeling in which we use a discrete representation obtained by neural nets trained in an unsupervised fashion. Equipped with this representation we can adapt recent powerful natural language processing models and demonstrate the ability to capture long-term dependencies in the signal as well as high-quality generations. Specifically, I will present our recent advancements in speech and audio language modeling involving different modalities for model conditioning and various application this can be used for.

Error correction code is a major part of the physical communication layer, ensuring the reliable transfer of data over noisy channels. Recently, neural decoders were shown to outperform classical decoding techniques. In this talk, we will present recent neural decoders based on Transformers and Diffusion models able to outperform existing neural decoders by very large margins.

AI is paving the way in academic research and innovation breakthroughs in the industry. Yet, much of the workload is cumbersome or scattered over siloed environments, making it hard to utilize and scale compute resources for Machine Learning development.

In this talk we'll discuss best practices for architecting, managing, and maintaining computing infrastructure for AI research in the next years to come.

Dynamic difficulty adjustment (DDA) is a process of automatically changing a game difficulty for the optimization of user experience. It is a vital part of almost any modern game. Most existing DDA approaches concentrate on the experience of a player without looking at the rest of the players. We propose a method that automatically optimizes user experience while taking into consideration other players and macro constraints imposed by the game. The method is based on deep neural network architecture that involves a count loss constraint that has zero gradients in most of its support. We suggest a method to optimize this loss function and provide theoretical analysis for its performance. Finally, we provide empirical results of an internal experiment that was done on 200, 000 players and was found to outperform the corresponding manual heuristics crafted by game design experts

ML models are not trustworthy often because they focus too much on "what" than "how". That is, they care only about whether they are solving the task at hand ("what") but not so much about solving it right ("how"). Having recognised this issue, the ML field has been shifting its focus from "what" to "how" for the last five years. Arguably, the most common approach to address "how" is to extend the familiar *benchmarking approach* that used to work well for the "what" phase: build a benchmark dataset and perform "fair" comparisons by fixing the allowed ingredients. This encourages more and more complex tricks that are likely to simply overfit to the given benchmark (e.g. ImageNet). However, for the "how" problem, I believe it is more important to transfer the way humans recognise the world (human explanations) into the computational domain and eventually utilise it for model training. I will give an overview of my previous search for such ingredients that make models more explainable and more robust to distribution shifts. I will then discuss exciting future sources of such ingredients.

In this talk, we explore the application of various ML architectures including language models to the problem of predicting SQL Query cost, i.e., CPU time and wall time to execute a SQL query command given only the query text. We show how a publicly available pretrained language model with fine tuning and appropriate data preprocessing can achieve state of the art performance on this highly valuable task. And we discuss a scalable approach to using these models in a real time machine learning framework that can generate fresh predictions from continuously integrated data without retraining.

A great battle over the morality of deploying Autonomous Weapons Systems (AWS) is being waged around the globe. The ultimate moral concern surrounds human dignity: is a human being demeaned by a machine deciding and in fact taking their life?  In this talk I argue that human dignity has different definitions based on context.

In the past few years, the European Union recognized the need to regulate artificial intelligence as it is a technology whose main characteristics and functionalities will enable a complete transformation of our society. One of many reasons why artificial intelligence is an industry game changer is its ability to instantly provide personalized online experiences based on vast amounts of available data.
 

The novel introduction of Artificial Intelligence (AI) into in-vitro fertilization (IVF) technologies (AI(i)IVF) is, at the same time, medically promising and ethically disconcerting. The new possibilities and complexities generated by involving intelligent, powerful algorithms in human creation and reproduction, raise a host of ethical issues for IVF patients, practitioners – embryologists in particular, and society at large. This paper seeks to identify and analyze various, as yet unexplored ethical implications of AI(i)IVF. We first originally contemplate a suitable framework for the ethical discussion. We suggest that a combination of the long-standing bioethics and the currently evolving AI ethics (nuanced, yet considerably overlapping) sets of principles should be thus applied: human-centric AI ethics principles – applied to the technological architecture of the AI(i)IVF tool, and the patient-centric principles of bioethics – applied to personal and societal aspects of implementing AI(i)IVF. We then respectively consider selected ethical principles to analyze the ethical merits and caveats of implementing AI(i)IVF, particularly for embryo evaluation and selection, while consistently debating whether putting a higher threshold for AI(i)IVF ethicality, than the one set for traditional (non-AI) IVF, is indeed warranted or justified. 
 

As the predictive power of AI gradually replaces human decision makers, calls for a "right for explanation" of automated systems gain simultaneous support. Accordingly, the technological domain entrusted with executing this right is continuously sharpening the professional tool, often related to as "explainability" or "interpretability", to generate explanations for end-users and decision subjects. This discussion explores the rudimentary question: Does explainability benefit end users? Leveraging insights from the legal and ML domains, a multidisciplinary analysis uncovers key complexities in the quest to generate users' explainability, and critically questions the intuitive appeal of AI explanations as a means to assist users in an automated world. 

The importance of detecting the existence of disease-related biomarkers in medical imaging data is undoubted as in most cases it may ameliorate patient care. Training a vision model to accurately annotate medical imaging requires a large dataset of annotated medical imaging. However, the existence of annotated volumetric imaging datasets, such as, 3D magnetic resonance imaging (MRI) scans, 3D optical coherence tomography (OCT) scans, and ultrasound (US) videos, is very limited as their annotation is both time- and cost-prohibitive. Per-se 3D-based vision models are thus practically bounded by a performance glass ceiling. In this work we present SLIViT, a 2D-based deep-learning heuristic framework that leverages transfer learning and vision transformers to accurately detect disease-related biomarkers in (potentially any) 3D medical imaging, given a low-sized 3D dataset for training. To exemplify SLIViT’s generic utility for diagnosing disease-related biomarkers in volumetric medical imaging, we applied it to three different data modalities tackling both classification and regression problems. We first show SLIViT’s performance superiority compared to both state-of-the-art methods and junior clinicians in classifying four different Age-related Macular Degeneration biomarkers (ROC AUC > 0.8) in two independent 3D OCT datasets. We then show SLIViT’s overperformance in measuring heart ejection fraction in US videos and show that with only an eighth of the original training dataset it can reproduce current state-of-the-art achievements (R^2 = 0.5). Lastly, we challenge SLIViT with a 3D MRI dataset where the order of the slices within each volume is not recorded. We show that even under these settings, we can accurately (R^2 = 0.7) model proton density fat fraction, a well-known biomarker for nonalcoholic fatty liver disease.

We propose an algorithm for learning a conditional generative model of a molecule given a target.  Specifically, given a receptor molecule that one wishes to bind to, the conditional model generates candidate ligand molecules that may bind to it.  The distribution should be invariant to rigid body transformations that act \textit{jointly} on the ligand and the receptor; it should also be invariant to permutations of either the ligand or receptor atoms.  Our learning algorithm is based on a continuous normalizing flow.  We establish semi-equivariance conditions on the flow which guarantee the aforementioned invariance conditions on the conditional distribution.  We propose a graph neural network architecture which implements this flow, and which is designed to learn effectively despite the vast differences in size between the ligand and receptor.  We evaluate our method on the CrossDocked2020 dataset, attaining a significant improvement in binding affinity over competing methods.

Drug development process is a lengthy and expensive process that can take more than 10 years and cost billions of dollars. Despite the major efforts, the process will fail in more than 90% of the cases due to safety concerns.
In this work we present a different approach for the task of assessing drug safety. An alternative to the animal models widely used today. We tackle the task using a BIO-AI approach. The main ingredient is a 'patient-on-a-chip' platform. The platform acts both as a human body simulator and as a data generator. The platform is capable of performing millions of interactions between the patient-on-a-chip and potential drugs, the data generated using the platform is used to train multimodal deep learning models to predict drug safety. Trained models are used by big pharma companies to predict safety of new drug candidates, and to provide personal safety recommendations for using existing drugs on specific patients.

Deep learning has revolutionized the field of machine learning with ground-breaking applications in various domains, including image processing, text analysis, and speech recognition. Deep learning has also become ubiquitous in bioinformatics and genomics. In this short talk, I will describe one of our sweeping successes in applying deep neural networks to high-throughput genomics datasets, and how we interpret the trained neural networks to discover new biology.

The following two fields have barely interacted before: reinforcement learning and molecular evolution. Here we develop a reinforcement-learning algorithm to tackle the challenge of reconstructing phylogenetic trees, which are used to describe the evolutionary relationships among a set of genome sequences. When only a few dozen of sequences are analyzed, billions of alternative phylogenetic trees could potentially describe the evolutionary patterns, thus rendering the search for the tree that best describes the data algorithmically challenging. Thus, all current algorithms for phylogenetic tree reconstruction use various heuristics to make tree inference feasible. Although recent studies have shown the potential of harnessing AI-based methods to reconstruct phylogenetic trees, the question regarding the overall tree-search strategy remains open. In this study, we provide a novel paradigm for predicting the maximum-likelihood tree by using reinforcement-learning techniques to learn an optimal strategy for tree search. We demonstrate the use of the developed deep-Q-learning agent on a set of unseen empirical data, namely, on unseen RL environments defined by nucleotide MSAs of up to 20 sequences. Our results show that the likelihood scores of the inferred phylogenies is as accurate as those obtained from widely-used software. It thus establishes a proof-of-concept that it is beneficial to optimize a sequence of moves in the search-space, rather than only optimize the progress made in every single move. This suggests that an RL-based method provides a promising direction for phylogenetic reconstruction.

We propose a new method to probe the learning mechanism of Deep Neural Networks (DNN) by perturbing the system using Noise Injection Nodes (NINs).
These nodes inject uncorrelated noise via additional optimizable weights to existing feed-forward network architectures, without changing the optimization algorithm.
We find that the system displays distinct phases during training, dictated by the scale of injected noise.
We first derive expressions for the dynamics of the network and utilize a simple linear model as a test case. 
We find that in some cases, the evolution of the noise nodes is similar to that of the unperturbed loss, thus indicating the possibility of using NINs to learn more about the full system in the future. 

Dimensionality reduction has become an indispensable tool in the modern algorithmic toolbox, often being used as a preprocessing step in data analysis and machine learning pipelines. Rigorously analyzing quality of such methods is therefore of great importance for applied community as well as of fundamental theoretical interest.
In this talk I will focus on metric dimensionality reduction, the task of embedding high-dimensional datasets into a small number of dimensions and preserving the metric structure of the input.
A rich body of research in Math and CS has been devoted to developing reliable and efficient embedding methods, with most techniques designed to optimize the worst-case quality guarantees. That is, the objective is to approximate all the pairwise distances of the dataset points. Yet much of this research line has little relevance to practical settings. Indeed, in nearly all practical applications of metric dimensionality reduction the fundamental criterion for measuring quality is its average performance over all pairs. Preserving distances among most of the pairs in the data will usually suffice for a successful application.
Various average-case quality measures have been suggested and studied from a practical point of view, and many heuristics have been developed to optimize these notions. However, almost nothing is known in terms of rigorous analysis and absolute bounds on the quality of these methods.
In the talk, I will discuss some of the most common practically oriented average-case quality measures. I will then focus on dimensionality reduction for Euclidean datasets and present our theorem stating that the good old random projection method is, surprisingly, the best possible for these measures. For the general datasets, I will present the first approximation algorithm with provable guarantees for the average quality measurement criteria and will conclude with several experimental results showing a clear advantage of the random projection-based method over two popular heuristics, PCA and ISOMAP.

The lower the distortion of an estimator, the more the distribution of its outputs generally deviates from the distribution of the signals it attempts to estimate. This phenomenon, known as the perception-distortion tradeoff, has captured significant attention in signal restoration, where it implies that fidelity to ground truth signals comes at the expense of perceptual quality (e.g. deviation from statistics of natural images). However, despite the increasing popularity of performing comparisons on the perception-distortion plane, there remains an important open question: what is the minimal distortion that can be achieved under a given perception constraint? In our work, we derive a closed form expression for this distortion-perception (DP) function for the mean squared-error (MSE) distortion and the Wasserstein-2 perception index. We prove that the DP function is always quadratic, regardless of the underlying distribution. This stems from the fact that estimators on the DP curve form a geodesic in Wasserstein space. In the Gaussian setting, we further provide a closed form expression for such estimators. For general distributions, we show how these estimators can be constructed from the estimators at the two extremes of the tradeoff: The global MSE minimizer, and a minimizer of the MSE under a perfect perceptual quality constraint. The latter can be obtained as a stochastic transformation of the former. Our theoretical results shed light on several topics that are subject to much practical activity, such as diverse image restoration and traversing between fidelity and visual quality during test-time, without training different models.

The field of reinforcement learning has demonstrated remarkable results in recent years. One challenge in bringing these results into the real world is the high risk-sensitivity in many applications. In the lecture, we will discuss the algorithmic difficulties caused by risk-sensitivity in reinforcement learning, namely, sample inefficiency and blindness to successful strategies. We will present the novel CeSoR algorithm for risk averse reinforcement learning. In a driving problem, for example, CeSoR not only prevents all accidents, but also learns intuitive risk-averse driving policies.

The lecture is based on the paper "Efficient Risk-Averse Reinforcement Learning" (NeurIPS 2022).
 

Machine Learning models, in particular Deep Neural Networks (DNNs), have expensive design and development processes, requiring substantial resources to produce a high-quality model (e.g., data collection, long training time, high level of computing capabilities, etc.). This motivates the consideration of DNNs as Intellectual Property (IP), and an attractive target for actors with the incentive to steal the model in order to benefit from the work of its owners. 
In recent years, various methods were suggested to detect models obtained illegitimately from their owners, however they have been shown to fail to demonstrate satisfactory robustness against model extraction attacks.
In this talk, we present an adaptive watermarking framework for machine learning models, leveraging the unique behavior present in a protected model as a result of a unique random seed initialized during the model training. This watermark is used to detect extracted models, which inherit the same unique behavior, indicating an unauthorized usage of the protected model's intellectual property (IP).
First, we show how an initial seed for random number generation as part of model training produces distinct characteristics in the model's decision boundaries, which are inherited by extracted models and present in their decision boundaries, but aren't present in non-extracted models trained on the same data-set with a different seed.
Based on our findings, we suggest the Robust Adaptive Watermarking (RAW) Framework, which utilizes the unique behavior present in the protected and extracted models to generate a watermark key-set and verification model.
We show that the framework is robust to (1) unseen model extraction attacks, and (2) extracted models which undergo a blurring method (e.g., weight pruning).
We evaluate the framework's robustness against a naive attacker (unaware that the model is watermarked), and an informed attacker (who employs blurring strategies to remove watermarked behavior from an extracted model), and achieve AUC values exceeding 0.9.
Finally, we show that the framework is robust to model extraction attacks with different structure and/or architecture than the protected model.
 

The hype of autonomous cars is getting real. Robots in the wild is a tough task. The main issue is an interaction with diverse and complex enviroment.  The evolutions steps are from fixed enviroments (like most robotics to today) to a more flexible less constrained but still somewhat limited in complexity. Also, the reliability is an important parameter. Can the robot do misstakes? 
In this talk we will try to outline the directions of the Robotics.

We see a plethora of deep learning methods, seeking to automatically infer a dense depth image from a single-color input image. This is important if we want to complement LIDAR sensors used in self-driving cars or to enable single-photo applications like image-editing.

Most State-of-the-Art applications these days rely on self-supervision to train deep neural networks to predict high quality depth-from-color.

Although showing promising results, these often do not translate into real world applications that easily. Either because of differences between synthetic data and real data domains or due to the scarcity of real data, limited in size and scale.

In this session, I am going to talk about the journey we had undertaken to develop a system that reconstructs a 3D model of a colon from an optical colonoscopy video for the purpose of increasing polyps' detection rate and raise standard of care.
 

With the recent success of generative models like chatGPT and Stable Diffusion it is clear that larger and larger datasets are needed to train those models. Billion or more image datasets like LAION are becoming more popular. With the growth of dataset size so does the challenge to curate and understand the content of those datasets. In this talk I present fastdup, our popular free GitHub tool that is the first tool (to the best of our knowledge) to scale to a billion image dataset. Using fastdup we have analyzed LAION 1B image dataset and share some surprising results. We plan to share a free version of cleaned LAION as a contribution to the computer vision community. Joint work with Dr. Danny Bickson and Prof. Carlos Guetrin from Stanford.

In today's digital era, dynamic business monitoring systems that continuously check for shifts in Key Performance Indicators (KPIs), data, and models are now a necessity. While many algorithms have been suggested for drift detection tasks, they often overlook some of the practical issues of monitoring. In the following talk, we will discuss three difficulties commonly encountered in monitoring tasks and some possible solutions to address them:
1.        Detection of model drift without access to outcome data - In some scenarios, there is no longer access to the dataset used for the training of the model. Without access to the original responses, it is not possible to quantify the expected variability in common model evaluation metrics (such as accuracy, precision, etc.). We suggest testing for model drift using a likelihood test, training a new model using the predictions as offsets, and testing whether the prediction improved significantly.
2.        Cumulative response over time - When monitoring certain attributes, such as auto insurance policy renewals, the actual response of whether the policy was renewed or not arrives with a delay; this can typically take up to 30 days. Therefore, it is difficult to make a direct comparison between the prediction and actual outcomes. To detect model drift, we suggest a test of the cumulative response over time.

3.        Alarm fatigue - Data drift alarm fatigue is a common problem when monitoring data for changes in the distributions of variables. Statistical tests for distribution equality can generate many false alarms as the number of observations increases, leading to alarm fatigue. To address this problem, we propose a three-fold solution based on multiple comparisons adjustments, equivalence testing, and setting business-relevant thresholds using simulations.
 

This is a talk about our paper: “BRUCE: Bundle Recommendation Using Contextualized item Embeddings” that was just published at the RecSys22 conference

In this paper, we adapted NLP Transformers to solve the bundle recommendation task, without any textual data.

Transformers’ success emphasized the importance of the relations between the sentence's words to the meaning of the whole sentence. A bundle is a set of items sold together, we believe that the relations between a bundle’s items are meaningful as well.

In this talk, I will present how we adapted Transformers for modeling bundles, their items, and most importantly - the user preferences towards the two.

At the AI Week conference, Yana will present lessons learnt from an AI Recommendation Product that enables Playtika to personalize game experience.

Recommendation algorithms usually aim to provide the best experience for a given user. However, in certain use cases, naive implementations may result in a greedy behavior which eventually leads to suboptimal performance of the model. In this talk, we will present our modeling approach that enables us to serve both sides of Ziprecruiter’s employment marketplace while recommending the most relevant jobs for job-seekers and the most relevant job-seekers for employers. Specifically, given a job-seeker’s contextual and behavioral data, our framework learns a jobseeker representation that induces a similarity metric between job-seekers and jobs representations, indicating a tendency to apply, by directly optimizing a joint multimodal embedding space using a multi task contrastive loss. Furthermore, the fact that job-seekers embeddings and the jobs embeddings reside on dual latent spaces enabling zero-shot transfer of the model to downstream tasks powering products for both sides of the marketplace. Finally, we will also discuss other implementation considerations as well as relevant business applications.

Mitigating cold-start situations is a fundamental problem in almost any recommender system. In real-life, large-scale production systems, the challenge of optimizing the cold-start strategy is even greater.
We present an end-to-end framework for evaluating and comparing different cold-start strategies. By applying this framework in Outbrain’s recommender system, we were able to reduce our cold-start costs by half, while supporting both offline and online settings. Our framework solves the pain of benchmarking numerous cold-start techniques using surrogate accuracy metrics on offline datasets – coupled with an extensive, cost-controlled online A/B test.
In my talk, I’ll start with a short introduction to the cold-start challenge in recommender systems. Next, I will explain the motivation for a framework for cold-start techniques. I will then describe – step by step – how we used the framework to cut the size of our exploration data by more than 50%.
 

What if you were able to know what are the most frequent topics our guests and partners are talking about? Get themed-based reviews? Understand the most important facilities for our guests? Have the most significant attributes and endorsements of a property or a destination? Detecting fraudulent messages? And many more..
In this talk, we will present how we build a multilingual multi-label topic classification model that supports zero-shot, with multiple architectures, using cross-encoder and bi-encoder.
We will show how fine-tuning BERT-like models on the tourism domain with a small dataset, can outperform other pre-trained models.

In recent years, factorization machine models (most notably FM and FFM) have dominated the realm of CTR prediction due to their ability to extract good user and item representations in sparse setups. While FFMs can produce more specific interactions between the different features, FMs are smaller models and are therefore the better choice for extremely sparse problems. There are, however, solutions that aim to get the best of both worlds. In this talk we will cover the basics of FM models, and present a hybrid variant which we used to better fit our production needs without loss of online performance. Finally, we’ll see how this solution can be integrated with alternative approaches from current literature.