Text-to-image diffusion can produce fantastic images, but it is often hard to control them using text prompts. They incorrectly match objects to their descriptions, cannot generate images of rare objects or multiple images of the same character, and cannot even count correctly. I will show how these issues can be corrected at inference time by smartly steering the attention components of the diffusion models.  
 

In the world of building games creating art assets is an extensive work. Gen AI can be a game changer of efficiency to this process, but you have to do it right.

In this lecture we will focus on the process of implementing Gen AI successfully and bring it from an idea to reality.

In control and optimization of critical systems (e.g. systems in healthcare, manufacturing and transportation), trial and error is typically prohibitively costly, prohibitively dangerous, or both.  A natural approach for learning to control and optimize without trial and error is using pre-recorded data for offline system identification --- i.e. for offline learning of a system model --- and then using this model for offline learning of an optimal controller.  This approach, namely offline system identification and optimal control, is achieving breakthrough success when implemented with overparameterized models (e.g. neural networks) trained by gradient descent (GD).  The success is fueled by implicit biases of GD, which yield not only in-distribution generalization, but also out-of-distribution generalization.  Towards elucidating this phenomenon, I will present a series of works that theoretically analyze implicit biases of GD when applied to overparameterized linear models in offline system identification and optimal control.  The results I will present offer theoretical explanations for the success of GD in control and optimization of critical systems, and suggest potential avenues for enhancing this success.

Entropy stands as a foundational concept with far-reaching implications, serving as a versatile quantitative metric for gauging information and uncertainty. This talk delves into the pivotal role of entropy in contemporary deep learning. While Shannon entropy has found enduring utility across numerous applications, its limitations come to the fore in scenarios involving the investigation of non-probabilistic or overly confident probability measures, such as deep neural networks. We introduce the concept of functional entropy, a perspective that departs from the conventional probability-based entropy and provide a comprehensive groundwork for functional entropy, encompassing key definitions for both conditional and marginal variants. We establish a decomposition property, underscoring its suitability for probing complex systems including the information and explanations of deep networks decisions. 

Generalization is a cornerstone concept in learning. So far, though, Generatlization Theory has mostly focused on discriminative tasks, where the concept of generalization is overall well understood. However, in recent years, generative models are beginning to take a more significant role in learning.  Unlike discriminative models, the criteria for generalization in generative models are not well-defined. 
In real-world applications, though, we need to decide when a model generalizes to new concepts and when it is copying the data  
 
In this talk, we will outline recent theoretical and empirical approaches to tackle this challenge, aiming to provide a rigorous framework for studying generalization in generative models and exploring the application of this theory to the issue of copyright infringement using AI models.  
 

Despite extensive research in the last decade, the gap between theory and practice seems to only grow. Can this trend be reversed? I will discuss a few research directions that try to lead to a closer interplay between deep learning theory and practice.

Fraud Fighting Experts Gilit Saporta and Yuval Rubin expose the impact of AI-powered fraud schemes on digital advertising. Focusing on bots that “listen” to audio/video streams, online radio and podcasts, we’ll learn how fraudsters generate fake traffic. By mimicking human listening patterns, these sophisticated bots blend in with legitimate traffic, which means that advertisers end up losing millions of dollars due to ad fraud schemes (to learn more, read about “CycloneBot” and “FM-Scam”, discovered by the fraud lab at DoubleVerify).

In this session, we explore the pivotal role of AI in transforming cybersecurity strategies. We examine why AI integration is essential in today's threat landscape.
We will highlight its ability to augment human capabilities, explore the symbiotic relationship between cybersecurity and AI, and analyse the transformative potential of AI, balancing its opportunities for innovation with ethical and privacy considerations and what is the role of leadership in this transformation.
 

In the rapidly evolving landscape of cyber threats, the integration of artificial intelligence (AI) technologies marks a significant paradigm shift in how we approach cyber security. This presentation delves into the emergence of AI as a cornerstone in the next generation of cyber security solutions. As traditional security mechanisms struggle to keep pace with increasingly sophisticated attacks, AI presents a promising avenue for bolstering our cyber resilience.

Attention-Deficit/Hyperactivity Disorder (ADHD) is prevalent and significantly impacts individuals, yet there is a lack of objective methods to evaluate the efficacy of medications. Consequently, obtaining effective treatment can be a challenging, long  and costly process. At iFocus Health, we address this issue by monitoring reading patterns and utilizing a machine learning algorithm to predict patient progress. This approach involves tracking reading patterns via a user’s webcam during a real-life reading task. The simplicity and relevance of this task to daily challenges allow it to be conducted at home, providing an objective and consistent measure. This measure can be employed by both patients and clinicians to identify the most suitable medication throughout the ADHD treatment journey.

We employ a Large Language Model (LLM) to convert unstructured psychological interviews into structured questionnaires spanning various psychiatric and personality domains. The LLM is prompted to answer these questionnaires by impersonating the interviewee. The obtained answers are coded as features, which are used to predict standardized psychiatric measures of depression (PHQ-8) and PTSD (PCL-C), using a Random Forest regressor. Our approach is shown to enhance diagnostic accuracy compared to multiple baselines. It thus establishes a novel framework for interpreting unstructured psychological interviews, bridging the gap between narrative-driven and data-driven approaches for mental health assessment.
 

In this work we present NOVA, an AI-driven approach for organellomic studies in neurons. We employ deep variational autoencoder architecture, in a self-supervised strategy, to systematically encode subcellular localization and morphology patterns from immunofluorescent images of 24 membrane bound and membrane-less organelles. NOVA provides an unprecedented depth to organellar microscopic studies and an integrative view of cell biology.

In this presentation, we unveil a groundbreaking approach to glaucoma detection that has the potential to revolutionize global eye health.
Glaucoma, a leading cause of blindness worldwide, often goes undiagnosed, particularly in low-income regions where traditional diagnostic methods are prohibitively expensive.
Our research introduces a novel, affordable, and precise method for intraocular pressure (IOP) monitoring, crucial for early glaucoma detection. By utilizing deep learning techniques, we examine the speckle patterns that the eye sclera reflects when sound waves stimulate it, enabling remote IOP measurement.
Through testing on pig eyes, our method demonstrates high accuracy and holds promise for global use.
This presentation showcases the transformative impact of AI-powered IOP monitoring in advancing glaucoma diagnosis and management, particularly in resource-constrained settings, and discusses the roadmap for future implementation and human trials.

Over 2 billion people worldwide rely on contaminated drinking water, leading to widespread waterborne diseases and the daily deaths of over 800 children under five from diarrheal diseases. One of the biggest problems that indirectly lead to waterborne diseases is the lack of measurements or the time delay between when a sample is taken and when accurate data relating to its contamination (or potential for contamination) can be reported. Our research explores the application of AI to water quality monitoring. It aims to develop models for real-time prediction of contaminants like fecal coliforms, E. coli, and fluoride, including developing AI-based sensor kits for effectively monitoring and reducing the impact of waterborne diseases.

Much of the research in reinforcement learning focuses on a single strategic agent optimizing its long term payoff. In this talk we will explore the possibilities of multiple agents to collaborate and improve their individual performance. We will also explore a related issue of delay when agents observe the outcomes with some delay, which is related to the communication between the agents. In this talk I will outline this research direction and focus on a few recent results both on cooperation and delay.

In multi-agent systems, goal-directed RL can be used not only to optimize the behavior of the ego agent but also to improve its understanding of other agents. This talk will first overview current goal-conditioned RL and then show how this problem formulation can be leveraged to improve reasoning and inference. It will conclude with a general discussion about goal-condition RL and its potential uses.

In the talk I will present several recent works on applications of machine learning in control theory - the study of manipulating a system's inputs to achieve a desired output.
I will focus on two key areas: adaptive control, where the controller learns and adapts in real-time, and online control, an emerging paradigm designed for dynamic and uncertain environments. 

Our results lie in the unique intersection between online learning and reinforcement learning, and control theory, and are enabled by numerous recent advances in these fields.

Contracts are the economic framework which allows a principal to delegate a task to an agent - despite misaligned interests, and even without directly observing the agent's actions. In many modern reinforcement learning settings, self-interested agents learn to perform a multi-stage task delegated to them by a principal. We explore the significant potential of utilizing contracts to incentivize the agents. We model the delegated task as an MDP, and study a stochastic game between the principal and agent where the principal learns what contracts to use, and the agent learns an MDP policy in response. We present a learning-based algorithm for optimizing the principal's contracts, which provably converges to the subgame-perfect equilibrium of the principal-agent game. A deep RL implementation allows us to apply our method to very large MDPs  with unknown transition dynamics. We extend our approach to multiple agents, and demonstrate its relevance to resolving a canonical sequential social dilemma with minimal intervention to agent rewards. Joint work with Dima Ivanov, Paul Duetting, Tonghan Wang and David Parkes.

AlphaZero is a central algorithm used for planning (sequential decision-making) in complex problems. Understanding the decisions and reasoning process of the algorithm is an important task on its own. At each time point the AlphaZero algorithm creates a Monte-Carlo-Tree -Search (MCTS) holding most of the reasoning behind its final decision. We provide tools and methods to decipher it and make it available for the extraction of human understandable explanations. 

Diffusion models revolutionized image generation but are imperfect image editors i.e., they create attractive images but the edits do not always follow the user’s intent. In this talk, I will argue that the key for success lies in counterfactual supervision, training on a dataset containing the precise user edits without changing any other attributes. Concretely, we use these ideas to propose ObjectDrop for object removal and insertion. We first collect a counterfactual dataset, capturing a scene before and after removing a single object, while minimizing other changes. Surprisingly, 2500 such counterfactual pairs are sufficient to turn a pretrained diffusion model into a state-of-the-art object remover, not only removing objects but also their effects on the scene. However, not all tasks require small datasets, indeed, photorealistic object insertion requires more counterfactual pairs than we can practically collect. We thus propose bootstrap supervision; we use our object removal model, trained on a small counterfactual dataset, to synthesize a large-scale object insertion dataset. Our approach significantly outperforms prior methods in photorealistic object removal and insertion, particularly at modeling the effects of objects on the scene.

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

One of the outstanding problems with large language models (LLMs) is their tendency to generate factually incorrect text, commonly referred to as "hallucinations". In this talk, I will present two key insights into why hallucinations occur and how we can predict them, even before the model generates a response. First, I will demonstrate that hallucinations often arise as a 'fallback behavior' of LLMs upon uncertainty. Through a systematic analysis, we will see that hallucinations are an evolution of other forms of undesirable degenerated texts. Second, I will show that it is possible to estimate how knowledgeable a model is about a given subject from its internal representations, which can be utilized to predict the factuality of the model's response before it is generated.

In this talk, we present an innovative concept for enhancing captured image quality through verbal interaction with the camera.  We introduce a method to improve image reconstruction by integrating textual descriptions of the captured scenes into a text-guided diffusion model, which significantly enhances image denoising and reconstruction. We show the benefit of this approach for improving regular imaging with a standard off-the-shelf camera. Moreover, we demonstrate the advantage of the approach for lensless imaging. We showcase a prototype flat lensless camera that uses a pre-trained diffusion model with a control network, enabling high-quality image recovery. These two examples for two different imaging devices demonstrate that talking to your camera can revolutionize imaging, offering superior results compared to conventional methods.

I will present ongoing work on narrative analysis of Holocaust testimonies. Other than the moral importance of the project and its value for Holocaust studies, I will argue that testimonies present an interesting test case for automatic narrative analysis, due to their relative abundance (tens of thousands of such testimonies were collected), and the fairly confined domain that they cover. I will briefly describe several lines of work we conduct on a dataset of 1000 testimonies collected by the USC Shoah Foundation, including identifying the progression of topics that a testimony addresses, the locations of its described events, and the religious disposition of the witness. I will conclude by discussing the higher-level goal of this project, which seeks to uncover recurring patterns in the type and organization of events in a large body of narrative documents used
by speakers when coming to tell a particular story.

Joint work with Yossi Adi, Avishai Elmakies, Maxim Ifergan, Renana Keydar, Amit Pinchevsky, Esther Shizgal, Itamar Trainin and Eitan Wagner 

We will dive into the advancements and changes in video generative models over the past year, exploring the latest developments and trends we have been witnessing at Lightricks while training such models. We'll learn how to effectively train and utilize video generative models in creative scenarios, unlocking the potential for innovative and artistic uses of this powerful technology.

How can we understand the core interests and passions of our buyers? How can we use this knowledge to introduce them to products they haven't yet discovered?
Our goal in this project is to address these challenges. On eBay, a destination for those seeking the unique and the unusual, buyers often arrive with specific passions, such as adding to a cherished collection. eBay's vast and varied inventory covers an extensive range of products and categories, providing a fertile ground for discerning these passions and delighting buyers with unexpected finds.
While typical recommendations tend to focus on items similar to those already viewed, our objective is to identify the long-standing passions that drive buyers' shopping habits, enabling us to suggest related products from different categories. 
The explosion of Large Language Models (LLMs) sent this project in unexpected directions, revealing novel ways to represent our buyers' interests.
 

EMNLP'23 paper (presented at EMNLP in Singapore in December 2023). Link: aclanthology.org/2023.findings-emnlp.877/ (I'll mainly present RQ3).

Concepts play a pivotal role in various human cognitive functions, including learning, reasoning and communication. However, there is very little work on endowing machines with the ability to form and reason with concepts. In particular, state-of-the-art large language models (LLMs) work at the level of tokens, not concepts. In this work, we analyze how well contemporary LLMs capture human concepts and their structure. We then discuss ways to develop concept-aware LLMs, taking place at different stages of the pipeline. We sketch a method for pretraining LLMs using concepts, and also explore the simpler approach that uses the output of existing LLMs. Despite its simplicity, our proof-of-concept is shown to better match human intuition, as well as improve the robustness of predictions. These preliminary results underscore the promise of concept-aware LLMs.

AI Sports Commentator is a case study of a domain-expert LLM-based application that replaces a human commentator in a game highlight video.
Sports commentators use a vast range of domain expertise. They are expected to integrate their knowledge and understanding of the game and the ever-changing dynamics between teams and players.

The standard use of LLMs for generating fan-engaging commentary presents a challenge due to the lack of these deep domain specializations.
We present a structured enriched play-details prompt, along with dynamic fewshots and instructions. This approach allows us to incorporate domain expertise while tailoring the style of the generated commentary to the specific event. Furthermore, we present a method based on a chain of thought that grounds the LLM's output to the given input event details, reducing hallucinations and irrelevant outdated knowledge use.

As AI regulatory requirements intensify (i.e. EU AI Act, etc.) and AI/gen AI models become more complex and less transparent, the ability to implement Responsible AI at scale is crucial to enable companies to successfully deploy AI models.
This talk will provide insights into how Mastercard develop and implement Responsible AI practices, through a cross-functional approach and highly scalable process, encompassing various aspects of Responsible AI - with a focus for this talk on the fairness (bias testing and bias mitigation), inclusiveness and transparency pillars. The talk will provide concrete examples and highlight some of the trade-offs and challenges we have faced in our Responsible AI Journey at Mastercard.

Real-time state estimation of dynamical systems is a critical endeavor in autonomous systems. While the Kalman filter (KF) stands as a cornerstone for systems governed by fully known linear Gaussian state space (SS) models, practical scenarios often deviate from this ideal, presenting nonlinear dynamics and partial information. In this talk I will discuss how the Kalman filter can be augmented using deep learning into KalmanNet, a novel real-time state estimator designed to navigate nonlinear dynamics and incomplete knowledge. KalmanNet leverages data-driven learning to enhance its adaptability, preserving the efficiency and interpretability of the classic algorithm while implicitly capturing intricate system dynamics. The talk will review the qualitative and quantitative gains arising from such hybrid model-based/data-driven design, and review several candidate applications.

A fundamental challenge in spectrum management is classifying every transmission within the spectrum. With the rise of Software Defined Radios (SDRs), the number and variety of transmitting devices have rapidly increased, making it impractical to rely solely on predetermined classes for classification. In this talk, we will introduce Novel Class Discovery (NCD) as a method to address the issue of having many unknown classes, which far exceed the number of known classes. We will also review recent state-of-the-art methods for both NCD and the related field of Deep Clustering, and show how these methods can be applied to solve this problem.

Lip-to-speech involves generating a natural-sounding speech synchronized with a soundless video of a person talking. Despite recent advances, current methods still cannot produce high-quality speech with high levels of intelligibility for challenging and realistic datasets. This talk presents LipVoicer, a novel method that generates high-quality speech, even for in-the-wild and rich datasets, by incorporating the text modality. Given a silent video, we first predict the spoken text using a pre-trained lip-reading network. We then condition a diffusion model on the video and use the extracted text through a classifier-guidance mechanism where a pre-trained automatic speech recognition (ASR) serves as the classifier.
We demonstrate the effectiveness of LipVoicer through human evaluation, which shows that it produces more natural and synchronized speech signals than competing methods (demo page: https://lipvoicer.github.io).
The presented LipVoicer is a joint work of Yochai Yemini, Aviv Shamsian, Lior Bracha, Sharon Gannot, and Ethan Fetaya. 
 

Predicting and controlling indoor air pollution values is essential for maintaining a healthy work and living environment. A straightforward approach for this task is to learn a regression model based on data from outdoor pollutant sensors alongside ventilation system sensor data. However, research shows that this type of prediction is extremely sensitive to the specific building structure and cannot be easily generalized across different locations. In this talk, we will discuss different methods to overcome this problem and present a robust and scalable deep learning architecture that combines data from a cold start period, during which both outdoor and indoor pollution levels are measured, as a static context to encapsulate the site-specific information.

The proliferation of wireless devices is leading to an ever-increasingly crowded radio spectrum. For example, different wireless communication systems that often work at the same time will coexist at the same frequency bands. Thus, interference mitigation—an instance of the source separation problem—stands out as a key challenge for maintaining high reliability in such systems. In this talk, I will discuss the potential of integrating artificial intelligence, particularly neural networks, into traditional communication systems architectures to enhance the separation of radio frequency (RF) signals in single-channel scenarios. Using a particularly interesting example of a signal type ubiquitous in WiFi/5G networks, I will explain some of the nontrivial technical challenges involved. These, if overcome, hold the promise of unprecedented gains, as I will further demonstrate. I will then briefly present the “RF Challenge”, an MNIST-like challenge that we have recently released, with the aim of stimulating work on RF signals within the AI community.

The multiple sequence alignment (MSA) problem is a fundamental pillar in bioinformatics, comparative genomics, and phylogenetics. Alignments are used, for example, to track the spread of viruses and to reconstruct the tree of life. We developed BetaAlign, the first deep learning aligner, which substantially deviates from conventional algorithms of alignment computation. BetaAlign draws on natural language processing (NLP) techniques and trains transformers to map a set of unaligned biological sequences to an MSA. We show that our approach is highly accurate, comparable and sometimes better than state-of-the-art alignment tools. We characterize the performance of BetaAlign and the effect of various aspects on accuracy; for example, the size of the training data, the effect of different transformer architectures, and the effect of learning on a subspace of indel-model parameters (subspace learning). Our findings further uncover the potential of NLP-based approaches for sequence alignment, highlighting that AI-based methodologies can substantially challenge classic tasks in phylogenomics, evolution and bioinformatics.

Peptide binders are short proteins that bind to larger proteins and subsequently interfere with their function. Despite their various applications in basic research and drug discovery, the rational design of peptide binders remains a challenging and time-consuming process. Herein, I will present an integrative peptide design protocol based on a sequence generative model trained on native protein interaction partners of the target. The model reverse-engineers the key sequence motifs underlying binding and generates novel sequences by recombination of these motifs. We tested our protocol on Calcineurin, a phosphatase involved in the activation of the cellular immune response. Following the generation of candidate sequences and filtering via molecular docking and high-throughput binding assays, we found that 70% of the designed peptides successfully interfered with the activity of Calcineurin. Altogether, generative modeling is a promising strategy for discovering peptide binders.

This talk presents research from a paper accepted for publication in Briefings in Bioinformatics, building on my PhD work. It details how deep neural networks, leveraging drug-target and drug-drug interaction datasets, identified eravacycline—a traditionally antibacterial drug—as a promising candidate for repurposing in treating pancreatic ductal adenocarcinoma (PDAC). Our study ranked all approved drugs and spotlighted eravacycline based on its efficacy signals. Subsequent in vitro investigations on BxPC-3 pancreatic cancer cells demonstrated eravacycline's ability to inhibit cell proliferation and migration while inducing apoptosis. These findings not only underscore the utility of machine learning models in uncovering unexpected drug-disease relationships but also illuminate a novel pathway for transforming antibacterial agents into cancer therapeutics. This work paves the way for further exploration of machine learning in drug repurposing, with potential implications beyond oncology

As artificial intelligence continues to advance and shape our world, it is crucial to examine its geopolitical implications. This panel, featuring experts from the United States, Europe, and Asia, will delve into the complex interplay between AI and international relations. The discussion will explore how AI is influencing global power dynamics, economic competition, and strategic decision-making. 

Attendees will gain valuable insights into the current state of AI from a geopolitical perspective and explore potential future scenarios. The panel will also highlight the importance of cross-border collaboration and dialogue in shaping the responsible development and use of AI worldwide.
 

Recent advances in deep learning have significantly improved target detection in RGB and monochrome images. However, detecting extremely small or amorphous objects remains challenging, particularly when targets are concealed or camouflaged. In military scenarios, distinguishing real targets from dummy objects is critical.
We propose an innovative approach that combines artificial neural networks with multispectral sensors, utilizing spectral features alongside traditional spatial, geometrical, and textural ones. Frames captured by multispectral sensors are aligned into a spectral cube using an AI-based Optical Flow Neural Network. An AI segmentation task then identifies target materials and shapes within the spectral cube.
This method enhances detection of small, sub-pixel targets and objects obscured by camouflage or vegetation, leveraging their unique spectral signatures to improve accuracy and reliability in complex scenarios.
 

Curriculum Learning (CL) is a widely used machine learning strategy that improves the learner's performance by allowing it to order the training samples during learning, similarly to the way humans learn. In this work, we address statistical aspects of CL, and consider a parametric learning problem with a target task and multiple source tasks. While only the target parameter is of interest, sampling from a source task might be beneficial if they are less noisy than the target task, while their corresponding parameters are very close. The learner is restricted by the total number of samples, and can adaptively choose how to allocate samples to each of the models. We define a strong-oracle learner as an ideal learner, which allocates all its samples to the most effective model (either the target or one of the sources). We show that achieving its performance is too ambitious for a learning algorithm, and advocate a weak-oracle learner, as a more realistic benchmark for CL algorithms. We first develop an elimination-based learning algorithm, and determine conditions that allow it to match the weak-oracle learner. We then consider lower bounds via minimax lower bounds. We reveal a few challenges associated with defining informative classes of problem-instances, propose two bounds, and determine the conditions under which the performance weak-oracle learner is provably optimal. Joint work with Omer Cohen and Ron Meir.

It’s been 18 months since ChatGPT launched by OpenAI and kick started the Generative AI era. As the industry is slowly stepping into phase two of this revolution, organizations are rapidly moving beyond AI ideation into large scale applications – it’s clear that the industry is moving past the AI Model obsession, exploring new opportunities to unlock business challenges, focused on giving companies a competitive edge. This shift is creating amazing opportunities for startups and developers, much like the early days of cloud computing, mobile tech, and social networks—arguably even bigger.
 
In this session, we will cover the AI-as-a-Platform vision by Microsoft, focusing on the massive opportunities for startups to build a new ecosystem based on AI Extensions. Furthermore, we will explore how to automate complex business processes responsibly with plugins and APIs, build supporting reasoning mechanisms for AI Models using proprietary data and 3rd party integrations, leverage memory for hyper-personalization capabilities, and adapt AI Extensions based on user feedback. Join us to discover how to stay ahead in this dynamic field and capitalize on the emerging opportunities in Generative AI.
 

Recent studies show that instruction tuning and learning from human feedback improve the abilities of large language models (LMs) dramatically. While these tuning methods can make models generate high-quality text, we conjecture that more implicit cognitive biases may arise in these fine-tuned models. Our work provides evidence that these fine-tuned models exhibit biases that were absent or less pronounced in their pretrained predecessors. We examine the extent of this phenomenon in three cognitive biases - the decoy effect, the certainty effect, and the belief bias - all of which are known to influence human decision-making and reasoning. Our findings highlight the presence of these biases in various models, especially those that have undergone instruction tuning, such as Flan-T5, Mistral-Instruct, GPT3.5, and GPT4. This research constitutes a step toward comprehending cognitive biases in instruction-tuned LMs, which is crucial for the development of more reliable and unbiased language models.

Curriculum Learning (CL) is a widely used machine learning strategy that improves the learner's performance by allowing it to order the training samples during learning, similarly to the way humans learn. 
In this work, we address statistical aspects of CL, and consider a parametric learning problem with a target task and multiple source tasks. While only the target parameter is of interest, sampling from a source task might be beneficial if they are less noisy than the target task, while their corresponding parameters are very close. The learner is restricted by the total number of samples, and can adaptively choose how to allocate samples to each of the models. 
We define a strong-oracle learner as an ideal learner, which allocates all its samples to the most effective model (either the target or one of the sources). We show that achieving its performance is too ambitious for a learning algorithm, and advocate a weak-oracle learner, as a more realistic benchmark for CL algorithms. 
We first develop an elimination-based learning algorithm, and determine conditions that allow it to match the weak-oracle learner. We then consider lower bounds via minimax lower bounds.  We reveal a few challenges associated with defining informative classes of problem-instances, propose two bounds, and determine the conditions under which the performance weak-oracle learner is provably optimal.

Joint work with Nir Weinberger and Ron Meir. 

Large Language Models (LLMs) have revolutionized how we create, interact with, and consume digital content. Consequently, their open-source ecosystem is gaining popularity at an abnormal rate.
 But are early adopters adequately addressing their security aspects and associated risks?
 
 In this talk, I will delve into the security risks of the LLM open-source ecosystem, emphasizing the need for a security-first approach to their adoption. I will present the results of an analysis conducted using the OpenSSF Scorecard comparing the security posture of the 50 most popular LLM-based projects on GitHub to that of other widely-used open-source projects, including projects from the OpenSSF Critical Projects list, raising significant concerns regarding their maturity and security posture.
 
 I will discuss different types of risks, demonstrate how they can manifest, and offer mitigating strategies to address them.

Malicious javascript code poses various risks. For example, it can steal sensitive and personal information, inject code and adds, and modify legitimate websites. When a security researcher investigates JavaScript code, she can analyze it with various tools to determine if it is a malicious or not. However, when there are millions of code snippets from multiple sources, the researchers cannot look at all of them. To make the work of the researchers more efficient, we used Large Language Models (LLMs) in three steps of the investigation. First, we use the LLM embedding layer together with clustering and outlier detection methods to group similar samples and highlight unknown samples. Then we utilize the LLM embeddings as an input for a classification model that provides a label and severity score for each sample so the researcher will be able to prioritize her work. Finally, we ask the LLM to analyse and explain the JavaScript code to help in the investigation process.

The growing reliance on GPS-based navigation and the advent of autonomous vehicles (AVs) bring significant enhancements in efficiency and safety but also increase susceptibility to cyber threats like GPS spoofing. Traditional cybersecurity measures often fall short in detecting and neutralizing these sophisticated attacks, which are designed to closely mimic legitimate GPS signals. To combat this, we propose the use of Deep Neural Networks (DNNs) within an unsupervised learning framework, aimed at improving the detection and mitigation of GPS spoofing. This approach excels in identifying subtle irregularities that diverge from established patterns, thereby offering a formidable defense against various spoofing strategies. Additionally, the study examines the implications of DNN inference speed on the ongoing operation and safety of AVs, ensuring rapid threat detection and response to maintain high safety standards.

Organizations have vast amounts of knowledge dispersed across many services, resulting in limited visibility - and this is particularly true with domain expertise such as security. 

A digital thread offers an integrated approach to combine disparate data sources across enterprise systems to drive traceability, visibility, and collaboration.

In this talk, you will learn how to create your own intelligent thread for your Engineering organization, using a combination of a Knowledge Graph built with Amazon Neptune using data from the Product Development Lifecycle and GenAI technologies using Amazon Bedrock in order to improve your product security. 

This session will showcase novel concepts like combining knowledge graphs and LLMs, and applying that to security. It will include a full demo and the associated project will be released as open source so that all the attendees will be able to experiment with it.