What are quantum computers, and why should I be interested in them? 

Will they have machine learning applications, soon, or ever?   

And what is entanglement, anyway? I will try to partially answer these questions and raise a few more in this very introductory talk, which will assume almost nothing except for some lack of common sense. 

The advances in deep learning algorithms together with the continuous increase in digital information volume provide a unique opportunity in the healthcare domain. Recent deep Learning studies have established remarkable performance in complex diagnostic tasks.  Nevertheless, some of the challenges include the need for interpretability to ease the human-machine collaboration, together with integration of imaging data and other types of data such as text, which is complex, noisy, heterogeneous and sometimes missing. In this talk we discuss several key applications spanning neuro CT, Chest X-ray and Spinal analysis with CT and demonstrate the results scaling large multi-domain dataset.

Deep learning has recently become an abundant technology for analyzing video data. However, the increasing resolution and frame rates of videos make real-time analysis a remarkably challenging task.

We present an overview of a novel architecture based on Redis, Docker, and TensorFlow that enables real-time analysis of high-resolution streaming video. The solution can serve advanced deep learning algorithms at subsecond rates and appears fully synchronous to the user despite containing an asynchronous backend. The talk offers a demo using visual inspection and shares results that highlight the solution’s applicability to real-time neural network processing of videos. The approach is generalizable and can be applied to diverse domains that require video analytics.

The talk will focus and the recent advances in Generative Adversarial Networks (GANs) for Image to Image Translations. I will begin by describing the problem and the distinction between supervised and unsupervised approaches, focusing on the approaches of pix2pix, DTN, CycleGAN, Unit and DistanceGAN. I will then describe recent approaches which allow for variation in the target domain as well as "one-shot"-translation. Lastly, I will talk about the application of these techniques in other domains such as audio and text. 

Who hasn't gotten an image or video with some dull skies taking up most of the space?

Professionals may edit these videos using advanced and time-consuming tools, to replace the sky with a more expressive or imaginative sky. 

In this work, we propose an algorithm for automatic replacement of the sky region in a video with a different sky.

The method is fast, achieving close to real-time performance on mobile devices and can be fully automatic.

Joint work with Tavi Halperin, Harel Cain and Michael Werman. To be presented at Eurographics 2019.

What does it mean to understand a visual scene? Deep models of visual recognition successfully exploit statistical regularities to recognize and localize objects relations and actions, but there is strong evidence that their "understanding" is limited.  I will describe a series of studies, both from our lab and from Nvidia research, aimed at learning intermediate representations that can be operated on. This allows models to generalize to new categories 

Quantum mechanics, which was born at the beginning of the 20th century and revolutionized our understanding of nature, implies that nature is far more counter intuitive and far richer than anyone could imagine. In the 1980’s, physicists understood that this richness of nature can be used to construct new computers that can outperform any classical computer. In the decades that followed, physicists have made huge progress in designing and controlling quantum systems, and today quantum computers seem to be closer than ever. What are the basic principles of quantum computers? how do they allow them to beat classical computers? What problems will they be best for? 

Deep sparse auto-encoders with mixed structure regularization (MSR) in addition to explicit sparsity regularization term and stochastic corruption of the input data with Gaussian noise have the potential to improve unsupervised abnormality detection. Unsupervised abnormality detection based on identifying outliers using deep sparse auto-encoders is a very appealing approach for medical computer aided detection systems as it requires only healthy data for training rather than expert annotated abnormality. 

In the task of detecting coronary artery disease from Coronary Computed Tomography Angiography (CCTA), our results suggests that the MSR has the potential to improve overall performance by 20-30% compared to deep sparse and denoising auto-encoders.

Manual network architecture search and hyper-parameter tuning of neural network based algorithms could be a very interesting task but also arduous. In this talk, we will present two novel methods developed in our lab for automating the network architecture search. With these methods, we are able to reach SotA results on various classification datasets within 1 GPU day per dataset. Are we heading towards a "CV as a Service" era?

Autonomous driving has come a long way albeit it's success is limited only to those areas where a High-Definition map is pre-built and known. Inspired by the recent achievements in Artificial Intelligence particularly of methods that combine trees and DNNs (e.g. Alpha Go Zero), this talk demonstrates how to effectively combine Deep Learning and convention path planning to drive in unknown areas. 

Common problems in the process of developing AI solutions for the medical field are highly unbalanced datasets on one hand and limited annotation resources on the other hand.

The use of Active Learning can dramatically help with both issues.

The task of Cell Detection is very important in digital pathology. For example, analyzing the quantity and density of immune cells can provide important indications on the progress of cancer.

This is a tedious task when manually done by pathologists and thus, automating this process is desirable.

Automating cell detection requires annotating large amounts of data, which is usually very unbalanced.

The DeePathology.ai Cell Detection Studio is a do it yourself tool for pathologists to train deep learning cell detection algorithms on their own data.

Using this tool, deep learning cell detection solutions can be easily created by the pathologist very quickly.

In the talk we will use the example of the DeePathology.ai Cell Detection Studio to demonstrate how Active Learning can be used for medical imaging annotation.

We will also present our approach for using active learning with unbalanced datasets.

Deep Neural Networks (DNNs) are usually executed by commodity hardware, mostly FPGA and GPU platforms, and accelerators, such as Google's TPU. However, when executing DNN algorithms, the conventional von Neumann architectures, where the memory and computation are separated, pose significant limitations on performance and energy efficiency, as DNN algorithms are compute and memory intensive.  Emerging memory technologies, known as memristors, enable in-place, highly parallel, and energy efficient analog multiply-accumulate operations. This is known as processing-near-memory (PNM). This talk will present the potential and opportunities of integrating memristors into DNN accelerators design.

Polygonal meshes provide an efficient representation for 3D shapes. They explicitly capture both shape surface and topology, and leverage non-uniformity to represent large flat regions as well as sharp, intricate features. This non-uniformity and irregularity, however, inhibits mesh analysis efforts using neural networks that combine convolution and pooling operations. In this paper, we utilize the unique properties of the mesh for a direct analysis of 3D shapes using MeshCNN, a convolutional neural network designed specifically for triangular meshes. Analogous to classic CNNs, MeshCNN combines specialized convolution and pooling layers that operate on the mesh edges, by leveraging their intrinsic geodesic connections. Convolutions are applied on edges and the four edges of their incident triangles, and pooling is applied via an edge collapse operation that retains surface topology, thereby, generating new mesh connectivity for the subsequent convolutions. MeshCNN learns which edges to collapse, thus forming a task-driven process where the network exposes and expands the important features while discarding the redundant ones. We demonstrate the effectiveness of our task-driven pooling on various learning tasks applied to 3D meshes.

The robustness of end-to-end driving policy models depends on having access to the largest possible training dataset, exposing the true diversity of the 10 trillion miles that humans drive every year in the real world. However, current approaches are limited to models trained using homogenous data from a small number of vehicles running in controlled environments or in simulation, which fail to perform adequately in real-world dangerous corner cases. Safe driving requires continuously resolving a long tail of those corner cases. The only possible way to train a robust driving policy model is therefore to continuously capture as many of these cases as possible. The capture of driving data is unfortunately constrained by the reduced compute capabilities of the devices running at the edge and the limited network connectivity to the cloud, making the task of building robust end-to-end driving policies very complex.

In this talk, I will give an overview of a network of connected devices deployed at the edge running deep learning models that continuously capture, select, and transfer to the cloud “interesting” monocular camera observations, vehicle motion, and driver actions. The collected data is used to train an end-to-end vehicle driving policy, which also guarantees that the information gain of the learned model is monotonically increasing, effectively becoming progressively more selective of the data captured by the edge devices as it walks down the tail of corner cases.

Learning to classify and localize instances of objects that belong to new categories, while training on just one or very few examples, is a long-standing challenge in modern computer vision. This problem is generally referred to as 'few-shot learning'. It is particularly challenging for modern deep-learning based methods, which tend to be notoriously hungry for training data.  In this talk I will cover several of our recent research papers offering advances on these problems using example synthesis (hallucination) and metric learning techniques and achieving state-of-the-art results on known and new few-shot benchmarks. In addition to covering the relatively well studied few-shot classification task, I will show how our approaches can address the yet under-studied few-shot localization and multi-label few-shot classification tasks. In addition to this talk, a detailed tutorial covering the few-shot learning field in general will be given by Dr. Joseph Shtok from my team.

Given Imaging revolutionized the GI world with the invention of Capsule Endoscopy 20 years ago. Capsule endoscopy was the first device able to visualize the Small Bowel and improved lives already of more than 3 million patients worldwide. The PillCam is the state of the art for Capsule endoscopy, owns the largest market share and continues to bring breakthrough innovations.

Today, the technology owned by Medtronic, keeps on developing additional applications for Small Bowel, Colon and Pan-enteric visualization for a variety of diagnosis and monitoring. The technologies use the most advanced AI, vision and machine learning algorithms to revolutionize once again this field and create the Intelligent Capsule Endoscopy.

In our presentation, we will present some of the unique technical achievements and challenges. Our deep learning based pathology detectors achieve expert human level performance with only a few hundreds of unique examples. Furthermore, this technology enables accurate localization of the GI tract without any need for additional inputs aside from the images.

Natural Stochastic Textures (NST) images exhibit self-similarity and Gaussianity that are the two main properties characteristic of Fractional Brownian Motion (fBm) processes. We consider non-pure NST images that contain also structural information. The latter is characterized by having profound local phase information, whereas the former is characterized by random spatial phase. In this meeting, we address primarily the fractal-based layer of the model and implementing it on the NST component. We also discuss applications where the approach can be used, with special emphasis on medical images. Examples of mammography and bone X-ray images are presented.

* Join work with Prof. Yehoshua Y. Zeevi

Modern day pharma organizations face mounting challenges in the field of drug research and development due to the pharma competitive landscape, the compelxity of treatment methods and increasing number of drug combination studies. At Roche we are persuing a Personalized Healthcare (PHC) approach to drug development, which is meant to match the best medicine to each specific patient. This approach has the potential to drastically improve drug efficacy for our patients but it also poses the challenge of patient-drug selection. The ammount of data available in the drug development process, from histology and clinical imaging, from genomics and sequencing, from real world data and from diagnostic methods present a unique oportunity for the extraction of novel insights and the development of new technologies. These novel machine learning techniques are helping us address these challenges and leverage this information to assist scientists in the research process, supporting descision making and bringing better drugs to patients, faster. 

The foundations of autonomous driving (AD) are based on advanced sensors and algorithms for environmental perception, especially for road and obstacles detection. The sensor set for AD is expected to contain a combination of low and high resolution, image and distance sensors, including cameras, Radars  and  Lidars. Low-level sensor fusion uses all sensor to generate a high-density pixel-level joint image-distance HD- RGBd  model through upsampling. The environmental perception is generated by processing this joint HD- RGBd  model through unified algorithms. While seems prone to loss of one sensor in the set, we actually developed built-in redundancy mechanisms that compensate for the loss of one sensor through alternative low-level fusion and detection of all the remaining sensors, with known loss of accuracy that can be used by the driving system to continue driving through at lower speed.

The eyes play a vital role in the perception of our evolving surrounding as well as in communication with other human beings. Apart from the behavioral and scientific research in medical diagnoses
and psychological studies, estimating gaze direction is also important in Mixed Reality, AI assistant devices, Automotive, Human-Computer and Robot interaction, User authentication and many other applications.

In this talk, I will give a short introduction to eye-tracking research, compare between the geometric (model based) and data-driven (appearance based) methods, with emphasis on the ad-vantages and disadvantages of deep-learning in this domain and present some interesting usages of this technology.

Media companies accumulate large archives and struggle to transform it into business value. Media is unstructured and therefore hard to manage at scale. Content categorization by topics is an intuitive approach that makes it easier for people to search. Many organizations turn to tagging their content manually, which is expensive and isn’t scalable. To automate this process, we apply multimodal topic detection and tracking (TDT). The information communicated through the speech, OCR, and identified celebrities is detected using Named Entity Recognition, clustered by concepts into similarity graphs, named using Wikipedia Categories ranker in an interpretable manner with evolvement over time.

Face swapping is one of the most important face synthesis problems. It has numerous practical application such as hairstyle replacement, face spoofing, and data augmentation for machine learning. One of the most prominent face swapping methods is DeepFakes, which has recently attracted a lot of attention from the media around the world by making face swapping more accessible to regular users. The rapid progress in face synthesis should raise concerns on its implications on society. What will happen when those methods become easily accessible and harder to distinguish from real images?

The road to fully autonomous vehicles has been filled with challenges. When will cars drive themselves? What will it take from a technological standpoint to get us there? During this presentation, we will discuss the industry and technological challenges surrounding self-driving vehicles. We will illuminate how these hurdles can be overcome by moving from a deep learning model into a completely new paradigm of machine learning technology – Autonomous AI

Augmented Reality is the combination of Computer Vision and Computer Graphics. As technology mature and is now being commoditized, new innovation can emerge. In this talk, we would share our story of Augmented Reality @ Trax. How we use it for multi-perspective geometry, indoor mapping and navigation, and even challenging the shopping experience.

Mobile photography has been transformed by software. While sensors and lens design have improved over time, the mobile phone industry relies increasingly on software to mitigate physical limits and the constraints imposed by industrial design. In this talk, I'll present the technology behind two recent projects we’ve developed for Google Pixel Phones: Synthetic Depth-of-Field with a Single-Camera (also known as Portrait Mode) and key algorithms for the recently released Night Sight mode.

Text detection and recognition systems have gained a significant amount of attention in recent years. Current state-of-art text detection algorithms tackle challenging text instances in natural images. In particular the problem of detecting multi-scale text in a single image still presents a challenge. A common paradigm for dealing with that challenge is simply, given a single-scale text detection algorithm, to re-run that algorithm on different rescaled versions of the original image. While this approach usually achieves a boost in results, it is wasteful and significantly increases runtime.

In our work, we present an approach that bypasses the need to re-run the same detection algorithm on multiple scales. We show that using a simple plug-and-play change in the architecture, we are able to transform a text segmentation Convolutional Neural Network to also detect text scales. Knowing the text scales allows us to adaptively re-scale text regions, and aggregate them into a compact image, which enables our network to detect the smaller text using only one additional pass. We present some qualitative and quantitative results on the ICDAR benchmark, showing that our approach offers a good trade-off between runtime and accuracy.

Viewing ultrasound imaging (US), or any other medical imaging modality for that matter, as an inverse problem, in which a latent image is reconstructed from a set of measurements, current research is focused mostly on learning the inverse operator producing an image from the measurements. The scope of our work differs sharply in the sense that we propose to learn the parameters of the forward model, specifically, the transmitted beams patterns (Tx), together with the receive beamforming (Rx). We demonstrate a significant improvement in the image quality compared to the standard patterns used in standard fast US acquisition settings.

Patch matching is a key ingredient in many image processing  and  computer  vision  applications.  Learning base approaches for patch-matching were shown to be successful, but were tailored towards specific tasks. A more general approach for improving the PM engine was recently introduced by Romano et. al, showing the potential improvement in using their modified similarity measure on several PM-based tasks. In our research, we investigate a deep-learning based strategy that preserve the order between distances and show that the order-preserving approach can improve upon previous work by a significant margin.

Multi-domain problems like grounding text in the Visual world are crucial for many real world applications, yet remain mostly unsolved.

We examine the effects of combining visual and linguistic representations revealing their fundamentally different nature yielding an imbalanced co-adaptation.

We introduce Cross Domain Normalization to dramatically stabilize learning, reduce overfit and speed up learning (up to 19X faster than Batch Normalization) though manipulation of cross domains statistics. With CDN our extremely simple model significantly outperforms all today’s SOTA models. The insights gained by investigating the linguistic and visual co-adapted parameters can be utilized for other multi-domain tasks and co-adaptation in general.

In traditional computer vision research, we were careful to strip and ignore the soundtrack. In the same manner, audio analysis traditionally ignored the available visual information in a video.
I will describe how where better scene understanding can be gained by using both sight and sound in video.
One example will show how speech can be enhanced when using a video showing the face of the speaker.
Also, video surveillance cases will be shown where the sound can enhance the understanding of scene activity.
Wide deployment of joint sight and sound analysis is limited by wiretapping and eavesdropping laws, preventing most surveillance and many wearable cameras from recording audio. Some possible approached to overcome these restrictions will be discussed.

Deep Learning has always been divided into two phases: Training and Inference. Deep networks are mostly used with large data-sets both under supervised (Classification, Regression etc.) or unsupervised (Autoencoders, GANs) regimes. Such networks are only applicable to the type of data they were trained for and do not exploit the internal statistics of a single datum.
We introduce Deep Internal Learning; We train a signal-specific network, we do it at test-time and on the test-input only, in an unsupervised manner (no label or ground-truth). In this regime, training is a part of the inference, no additional data or prior training is taking place. This is possible due to the fact that one single instance (be it image, video or audio) actually contains a lot of data when internal statistics are exploited. In a series of papers from the last year, that will be reviewed throughout the talk, I will demonstrate how we applied this framework for various challenges: Super-Resolution, Segmentation, Dehazing, Transparency-Separation, Watermark removal. I will also show how this approach can be incorporated into Generative Adversarial Networks by training a GAN on a single image for the challenge of retargeting.

Few-shot learning is a rapidly evolving field in computer vision, where the standard tasks of classification and detection are addressed in a challenging scenario with only few examples per class available for training. Using the approaches of meta-learning, metric learning, synthesis (hallucination) and data augmentation, substantial progress has been made in this field during the recent years. We will provide an overview of the state of the art in few-shot learning, highlighting and describing the main methods and results in the aforementioned directions.

The greatest challenge regarding simulation is finding the matrix that compares it to real life.

In this session, Danny will present a mathematical matrix that defines this relation and show how a proper simulation can be constructed based on deep learning techniques. He will also supply live examples of the Cognata simulation platform.

As video becomes the most popular form of communication and information sharing, automatic video comprehension is becoming a key requirement in many use cases. Fully understanding video poses many challenges, that are being addressed by a variety of AI techniques. In this talk, I will describe some of these challenges and present several ongoing projects in IBM Research AI in the space of Video Comprehension. These include segmentation of video into semantic scenes, recognizing objects and human actions and activities, identifying highlights in videos, few shot learning for object recognition. I will also present the Moments in Time public data set created by IBM Research AI, with its millions of video clips, to support the research and development of state of the art AI algorithms for the comprehension of subtle actions and activities in videos.

While a comprehensive theory of deep learning is still to be developed, one of the most interesting existing theoretical frameworks is the information bottleneck theory of deep neural networks.  In this framework we consider the information flow from the input layer through the layers, which form a cascade of filters of the irrelevant information about the desired label. The theory explains how stochastic gradient descent can achieve optimal internal representations, layer by layer, and what each layer represents. Moreover, it explains the Computational benefits of the hidden layers and the specific information encoded by the weights of each layer in the deep network.

Reinforcement learning (RL) is a branch of machine learning concerned with using experience gained through interacting with the world and evaluative feedback to improve a system's ability to make behavioral decisions [Littman, Nature 2015].  In the recent years RL achieved some major successes, which include: super-human GO playing, challenging robotic manipulation tasks execution, super-human computer game playing, and autonomous helicopter flight. This tutorial will start from RL's foundations, continue with some of RL's core techniques, and will cover some of RL's recent successes. This tutorial assumes no prior background in RL - general CS/math background will suffice.

We succeeded to train an agent to drive in a specific unvisited area using only a realistic 3D model of the area which was built using merely aerial images. The training system is end to end: the inputs are realistically rendered images from the 3D model to the agent perspective and the outputs are driving commands to the moving platform. While training, the system got only sparse weak supervision from collision detections with the 3D model. Our main novelty is the separate treatment to the two linked challenges: navigation and obstacle avoidance, which both affect the platform location and position, but have different rewards distribution and thus hard to learn simultaneously. The two are approached using different methods and combined in the final agent behavior. We hope to use the trained agent for driving a real vehicle inside the modeled area. 

Donde Search uses Computer Vision and Natural Language Processing to analyse product pages and extract meaningful information to improve navigation, merchandising, personalization, and search across e-commerce platforms.

In this talk we will present some of Donde's solutions and discuss their underlaying technologies. We will show how Donde's technology transforms a challenging and sometimes frustrating search experience into a natural and engaging visual process that allows humans to navigate their way in a database including hundreds of thousands of items. We will discuss the gap between state-of-the-art deep-learning algorithms and real working solutions for enterprises, including some of the most common problems in transforming working algorithms into product like handling structured data, combining texts and images and using unsupervised methods for data-augmentation and for allowing better training and learning.