Pure model-based approaches are today often insufficient for solving complex inverse problems in medical imaging. At the same time, we witness the tremendous success of data-based methodologies, in particular, deep neural networks for such problems. However, pure deep learning approaches often neglect known and valuable information from the modeling world, are prone to instabilities and are not interpretable.

In this talk, we will develop a conceptual approach by combining the model-based method of sparse regularization by shearlets with the data-driven method of deep learning. Our solvers are guided by a microlocal analysis viewpoint to pay particular attention to the singularity structures of the data. Finally, focussing on the inverse problem of (limited-angle) computed tomography, we will show that our algorithms significantly outperform previous methodologies, including methods entirely based on deep learning.

Although deep learning (DL) models have been shown to outperform other frameworks for a variety of medical contexts, inference in the presence of pathology in medical images presents challenges to popular networks. Errors in deterministic outputs lead to distrust by clinicians and hinders the adoption of DL methods in the clinic. Moreover, given that medical image analysis typically requires a sequence of inference tasks to be performed, this results in an accumulation of errors over the sequence of outputs. This talk will describe recent work exploring (MC-dropout) measures of uncertainty in tumour detection and segmentation models in patient images and illustrate how propagating uncertainties across cascaded medical imaging tasks (e.g. MR image synthesis) can improve DL inference. The models have been successfully applied to the MICCAI BRaTs brain tumour segmentation challenge dataset, and they were also used to devise metrics for ranking competing methods in the new BRaTs sub-challenge on Quantifying Uncertainties in Brain Tumour Segmentation.

Geometric deep learning has recently become one of the hottest topics in machine learning, with its particular instance, graph neural networks, being used in a broad spectrum of applications ranging from 3D computer vision and graphics to high energy physics and drug design. Despite the promise and a series of success stories of geometric deep learning methods, we have not witnessed so far anything close to the smashing success convolutional networks have had in computer vision. In this talk, I will outline my views on the possible reasons and how the field could progress in the next few years. 

Deep learning is experiencing unprecedented success in recent years, delivering state of the art performance in numerous application domains.  However, despite its extreme popularity and the vast attention it is receiving, this technology suffers from various limitations --- in terms of stability, reliability, explainability and more --- hindering its proliferation.  In this talk, I will argue that theoretical analyses of deep learning may assist in addressing such limitations, by providing principled tools for neural architecture and optimization algorithm design.  Two examples will be given: (i) application of tensor analysis and quantum mechanics for configuring the architecture of a convolutional neural network; and (ii) dynamical analysis of gradient descent over linear neural networks for enhancing convergence and generalization properties.

The deep fakes originally posted in autumn of 2017 were hardly the first face swap examples the computer vision and graphics communities have seen. These recent deep fakes did, however, introduce something new compared to earlier methods: photorealism. Naturally, as deep fakes become harder for us to distinguish from genuine photos, so does automatic deep fakes detection become a harder task for computer vision systems. In this talk I will present the Deep Fake Detection Challenge (DFDC) which was recently introduced by a consortium of industry partners as a means of promoting research into this deep fake detection task. I will explain why DFDC is unique and discuss some of the experience and consequences already obtained on this challenge. In particular, I will show evidence that generalizing deep fake detection to unknown face manipulation techniques is hard. I will then propose a novel approach for detecting such manipulations, which leverages well-known perceptual phenomena to provide a general, non subject specific, fake detection signal. 

In most videos we capture, humans constitute the salient objects in the scene. Thus, developing computational methods for analyzing, visualizing and manipulating people in videos plays an important role in many applications (e.g., Augmented Reality, Robotics). In this talk, I'll highlight a few of my recent works, each touching on a different aspect of people's geometry and motion in ordinary videos. More specifically, (i) a method that transforms an ordinary video of a person moving into a 3D sculpture representing its share and motion. (ii) a deep-learning based model that predicts the geometry (depth) of people in videos in the challenging case where both the camera and the people in the scene are freely moving, and (iii) a neural-rendering based model for retiming--speeding up, slowing down, or entirely freezing--certain people in videos, while automatically re-rendering properly all the scene elements that are related to those people, like shadows, reflections, and loose clothing. All these methods take just an ordinary video as input and learn concepts of natural motion, geometry and scene decomposition, without requiring any manual labels.  I'll demonstrate new video effects on various real-world complex videos such as dancing, groups running, or kids jumping on trampolines. 

Scene graphs are detailed semantic descriptions of images. In this talks I will describe methods for annotating images with scene graphs, learning how to annotate from weak supervision, and generating images from scene graphs. In particular, I will discuss questions of representation invariance in these architectures.

I will show how complex visual inference can be performed with Deep-Learning, in a totally unsupervised way, by training on a single image -- the test image itself. The strong recurrence of information inside a single image provides powerful internal examples, which suffice for self-supervision of CNNs, without any prior examples or training data. This gives rise to true “Zero-Shot Learning”. I will show the power of this approach to a variety of problems, including super-resolution, segmentation, transparency separation, dehazing, image-retargeting, and more.

I will further show how self-supervision can be used for “Mind-Reading” (reconstructing images from fMRI brain recordings), despite having only little training data.

One of the challenges facing researchers is having their research make it into a product and make business impact. They have to “shop” their research in the organization, to find the right partner who needs exactly the research they have done.

In this talk, we will discuss how we alleviate this issue in Alibaba’s DAMO Israel Machine Intelligence lab. We will show three specific examples: 1. How we can deploy best-in-class networks to Alibaba’s mobile apps, keeping them lightweight and efficient through state-of-the-art pruning and network architecture search; 2. How we handle image classification in the real world, with best-in-class multi-class labeling; 3. How we can provide real-world image editing - “object deletion” with a best-performing inpainting algorithm that is especially suited to the task at hand. 

These three are examples of how we increase the impact of our research, by first identifying the fundamental technologies required for the products we are involved in, understanding their requirements, and then executing world-leading research to provide a business edge.

Humans can drive a car using a vision-only system, without relying on 3D sensors at all, and achieve a remarkable high accuracy. Can we match this ability using computer vision? The talk will focus on some of the challenges, including machine learning with extremely high accuracy, lifting a 2D projection back to the 3D world, and developing decision-making algorithms that are robust to sensing errors.

Perception algorithms is a well studied and active research field. Despite numerous research breakthroughs in the field and industry high focus, productization of perception systems remains a challenge. This talk will discuss the complexity of creating perception algorithms for real products. This includes advanced sensor fusion technics, uncertainty creation, algorithm generalization to different sensor types and meeting aggressive compute constraints while keeping high accuracy. Additionally, since perception is a data hungry system, we will discuss how improvement over time is achieved as data becomes available.

More than eight years have passed since the “ImageNet moment” spurred a fundamental shift in the direction of information technology, followed by distinct waves of startups founded to commercialize the use of CNNs and other ANNs in practical settings.

While the first waves focused on their novel engineering capabilities and technical depth, the current wave is fully engaged in growing businesses based on Intelligent Edge technologies, that face a wide range of competitors from public and other startup companies.

Development in the underlying hardware technology enabling large advances in energy efficiency, accuracy, and latency of compute operating at the edge has begun to accelerate along with the emergence of the market, with key innovations in embedded memory, sensor technology, and the integration workflow of creating novel Intelligent Edge systems. In this talk we will cover some of the challenges that different categories of startups commercializing Intelligent Edge are facing, and how partnerships that leverage the depth of capabilities present in the semiconductor industry can be employed to differentiate and advantage startups in the space.

 

Single image super resolution (SR) has seen major performance leaps in recent years. However, existing methods do not allow exploring the infinitely many plausible reconstructions that might have given rise to the observed low-resolution (LR) image. These different explanations to the LR image may dramatically vary in their textures and fine details, and may often encode completely different semantic information. In this work, we introduce the task of explorable super resolution. We propose a framework comprising a graphical user interface with a neural network backend, allowing editing the SR output so as to explore the abundance of plausible HR explanations to the LR input. At the heart of our method is a novel module that can wrap any existing SR network, analytically guaranteeing that its SR outputs would precisely match the LR input, when downsampled. Besides its importance in our setting, this module is guaranteed to decrease the reconstruction error of any SR network it wraps, and can be used to cope with blur kernels that are different from the one the network was trained for. We illustrate our approach in a variety of use cases, ranging from medical imaging and forensics, to graphics.

In order to perform complex calculations in the field of artificial intelligence, a great deal of computational power is required, which is also able to handle information in very large volumes. In fact, the need for artificial intelligence computing and the need to solve increasingly complex computing problems are pushing the computing market forward consistently, including moving to GPU processing units and designing new components that will be specifically tailored for artificial intelligence computing. For Israel

The ability to stay up-to-date and relevant is required, along with the ability to research and innovate independently. It is important to emphasize that the needs are not only in the power of calculation itself, but also in storage, communication, support and more..

When we come to define the infrastructure needs for artificial intelligence uses, we need to ask two questions: The first is who our users are. The second is what their needs are regarding the following topics: access to shared information, cost savings, ability to solve large-scale problems, classification constraints, confidentiality and security, performing innovative hardware and software testing, community support, education and training.

We introduce a novel neural network-based partial differential equations solver for forward and inverse problems. The solver is grid free, mesh free and shape free, and the solution is approximated by a neural network.
We employ an unsupervised approach such that the input to the network is a points set in an arbitrary domain, and the output is the set of the corresponding function values. The network is trained to minimize deviations of the learned function from the PDE solution and satisfy the boundary conditions.
The resulting solution in turn is an explicit smooth differentiable function with a known analytical form.
Unlike other numerical methods such as finite differences and finite elements, the derivatives of the desired function can be analytically calculated to any order. This framework therefore, enables the solution of high order non-linear PDEs. The proposed algorithm is a unified formulation of both forward and inverse problems where the optimized loss function consists of few elements: fidelity terms of L2 and L infinity norms, boundary and initial conditions constraints, and additional regularizers. This setting is flexible in the sense that regularizers can be tailored to specific problems. We demonstrate our method on several free shape 2D second order systems with application to Electrical Impedance Tomography (EIT). 

Super-resolution (SR) methods typically assume that the low-resolution (LR) image was downscaled from the unknown high-resolution (HR) image by a fixed 'ideal' downscaling kernel (e.g. Bicubic downscaling). However, this is rarely the case in real LR images, in contrast to synthetically generated SR datasets. When the assumed downscaling kernel deviates from the true one, the performance of SR methods significantly deteriorates. This gave rise to Blind-SR - namely, SR when the downscaling kernel ("SR-kernel") is unknown. It was further shown that the true SR-kernel is the one that maximizes the recurrence of patches across scales of the LR image. In this paper we show how this powerful cross-scale recurrence property can be realized using Deep Internal Learning. We introduce "KernelGAN", an image-specific Internal-GAN, which trains solely on the LR test image at test time, and learns its internal distribution of patches. Its Generator is trained to produce a downscaled version of the LR test image, such that its Discriminator cannot distinguish between the patch distribution of the downscaled image, and the patch distribution of the original LR image. The Generator, once trained, constitutes the downscaling operation with the correct image-specific SR-kernel. KernelGAN is fully unsupervised, requires no training data other than the input image itself, and leads to state-of-the-art results in Blind-SR when plugged into existing SR algorithms.

In the computer vision industry, gathering and manually annotating data is the most substantial bottleneck in the development of deep learning solutions. A promising solution is to generate data through 3D simulations as they provide perfect annotations and densely sample edge cases that real datasets fail to capture. Yet, a known shortcoming of this method is the domain gap between the simulated and real world domains. We show it can be overcome through the mutual use of Photorealistic Simulation and Domain Adaptation. To validate our claim on a study case, we generated simulated datasets that achieve state-of-the-art performance for 2D hand joints estimation. In this talk, we will present this methodology as a base for solving practical computer vision challenges in a wide range of domains. 

Background models are widely used in computer vision. While successful Static-camera Background (SCB) models exist, Moving-camera Background (MCB) models are limited. Seemingly, there is a straightforward solution: 1) align the video frames; 2) learn an SCB model; 3) warp either original or previously-unseen frames toward the model. This approach, however, has drawbacks, especially when the accumulative camera motion is large and/or the video is long. Here we propose a purely-2D unsupervised modular method that systematically eliminates those issues. First, to estimate warps in the original video, we solve a joint-alignment problem while leveraging a certifiably-correct initialization. Next, we learn both multiple partially-overlapping local subspaces and how to predict alignments. Lastly, in test time, we warp a previously-unseen frame, based on the prediction, and project it on a subset of those subspaces to obtain a background/foreground separation. We show the method handles even large scenes with a relatively-free camera motion (provided the camera-to-scene distance does not change much) and that it not only yields State-of-the-Art results on the original video but also generalizes gracefully to previously-unseen videos of the same scene. The talk is based on [Chelly et al., CVPR '20]. This is joint work with Vlad Winter, Dor Litvak, Oren Freifed (all from BGU CS) and David Rosen (MIT).

As deep learning is showing potential value in different markets, there is an increasing need to be able to run inference efficiently on edge devices.

In this talk we will focus on the fundamental characteristics of deep learning algorithms, analyze the challenges they introduce to the classical 60 years old Von-Neuman processing approach and review the guidelines to building more efficient domain specific processing architecture.

Beginning with some theoretical reasoning behind domain-specific architectures and their implementation in the field of deep learning, and more specifically for machine vision applications. We will use various quantitative measures, and more detailed design examples in order to make a link between theory and practice.

Hailo has developed a specialized deep learning processor that delivers the performance of a data center-class computer to edge devices. Hailo’s AI microprocessor is the product of a rethinking of traditional computer architectures, enabling smart devices to perform sophisticated deep learning tasks such as imagery and sensory processing in real time with minimal power consumption, size and cost.

Network architecture search (NAS) achieves state-of-the-art results in various tasks such as classification and semantic segmentation. Recently, a reinforcement learning-based approach has been proposed for Generative Adversarial Networks (GANs) search. In this work, we propose an alternative strategy for GAN search by using a proxy task instead of common GAN training. Our method is called DEGAS (Differentiable Efficient GenerAtor Search), which focuses on efficiently finding the generator in the GAN. Our search algorithm is inspired by the differential architecture search strategy and the Global Latent Optimization (GLO) procedure. This leads to both an efficient and stable GAN search.  After the generator architecture is found, it can be plugged into any existing framework for GAN training.

Understanding 3D human pose from images is a challenging task, but with a large number of potential applications. In this talk, we will present our work on looking beyond the weak perspective assumptions, usually employed in the conventional approaches to 3D human pose estimation. Through experimentation on the 3D poses in the wild (3DPW) dataset, we show that using full perspective projection, with the correct camera center and an approximated focal length, provides favorable results over existing methods. Our algorithm has resulted in a winning entry for the 3DPW Challenge, reaching first place in joints orientation accuracy.

WSC Sports’ main goal is developing sports’ action detection and event recognition algorithms with a very high success rate, in order to automatically generate highlight videos from sports broadcasts. Due to its high recall rates, an object detector is a commonly used tool in our proprietary action detection algorithms

Its main limitation is its precision rate on a sports broadcast, where it gets “real world” images, which produce false positives.

In this presentation, we propose a method for improving object detection models’ precision, without enlarging the train set, by using preceding and successive convolutional neural networks.

Given a pair of images of the same geographic area taken at different times, we wish to detect changes between them. Change detection is a challenging task. It is required to distinguish between fundamental changes, often man made, and insignificant natural ones. The latter may result from changing lighting, weather, camera pose, slight vegetation movement due to wind, and small errors in image registration. We address the change detection problem by training a learned descriptor using registered image pairs. Our fully convolutional CNN-based descriptor can efficiently detect changes in large aerial image pairs. It is shown to generalize well for a completely new scene and type of changes, while being robust to registration errors. The labeling of each image pair as similar or different is implied by the automatic registration process. Therefore, no manual annotation of any kind is required. While the lack of supervision results in label noise, the algorithm proves highly robust to it.

Many real-world applications suffer from lack of ground-truth.  we propose innovative an end to end network, dealing with zero-shot or few-shot segmentation.

We will show an innovative visual intuition that makes triplet-loss post processing redundant and enables end-to-end networks for many applications.

In addition, our network has the advantage of dealing with noisy labeling, by letting the network optimize accuracy without compromising consistency.

We introduce the first completely unsupervised correspondence learning approach for deformable 3D shapes.

Key to our model is the understanding that natural deformations, such as changes in pose, approximately preserve the metric structure of the surface, yielding a natural criterion to drive the learning process toward distortion-minimizing predictions. On this basis, we overcome the need for an- notated data and replace it by a purely geometric criterion. The resulting learning model is class-agnostic, and is able to leverage any type of deformable geometric data for the training phase. In contrast to existing supervised approaches which specialize on the class seen at training time, we demonstrate stronger generalization as well as applicability to a variety of challenging settings. We showcase our method on a wide selection of correspondence benchmarks, where the proposed method outperforms other methods in terms of accuracy, generalization, and efficiency.

Today, there is an increasing desire by patients and hospital managers to make trainings for residents and surgeons before performing actual operations. This desire made modern surgery syllabus to include use of simulators for improving surgical skills in teaching programs. In this talk I will review our computerized algorithms aim to score the performance of laparoscopic cutting and suturing operation in both general surgery and ophthamology medical fields . Using our algorithms, human assessment is not necessary and the quality of the surgery outcomes is objectively evaluated. 

Tumor Treating Fields (TTFields) is an FDA approved treatment for specific types of cancer and significantly extends patients’ life. The intensity of the TTFields within the tumor is associated with the treatment outcomes: the larger the intensity, the longer the patients are likely to survive.  This requires optimizing TTFields transducer array location such that their intensity is maximized around the tumor. Finding the best array placement is a multi-stage computationally expensive optimization problem. Here, we present a novel method that incorporates machine learning and deep learning in order to allow physicians a better TTFields treatment planning.

In this talk we will discuss our recent advances in few-shot learning, a regime where only a handful of training examples (maybe just one) are available for learning novel categories unseen during training. We will cover a method for few-shot classification that is capable of matching and localizing instances of novel categories, despite being trained and used with only category level image labels and without any location supervision, also opening the door for weakly supervised few-shot detection. We will cover a method for meta-learning a model that automatically modifies its architecture to better adapt to novel few-shot tasks. Finally, we will discuss the limitation of the current few-shot learning methods when handling extreme cases of domain transfer, and offer a new benchmark and some ideas towards cross-domain few-shot learning.

We are at the forefront of a revolution in how we interact with the world. Brain-Machine Interfaces (BMI) is technology that uses neural signals and advanced algorithms to decode the brain, enabling interactions with devices using only your mind. Recent advances in BMI are based on progress in invasive and non-invasive sensors, and the marriage of neuroscience and machine learning. In the next few years, invasive BMI applications will restore movement to paralyzed patients, enable stroke patients to speak, and enable the blind to see. Non-invasive applications will change the way we interact with devices, digital experiences, entertainment, and mental healthcare.

3D metrology is a new fascinating field in the semiconductor industry. Shrinkage of planar devices has reached its physical limit and advanced nodes resort to 3D design to increase the feature density in the device. Reliable measurements of these 3D structures are crucial for a chip development process.

We propose a novel supervised ML (Machine Learning) based solution for inferring 3D structure from 2D SEM (Scanning Electron Microscope) images. Our algorithm reached sub-nanometer accuracy and high precision.

The generality of our method and its ability to extract hidden information from SEM images open the door to a plethora of applications in 3D metrology for memory and logic devices.

There is a growing interest in person re-identification (ReID) but most studies can not be used for real-world applications.

Academic leading deep learning ReID models remain large-sized and computationally expensive, therefore unfavourable for real deployments in scalability, and specifically, their architectures are not suited for running on the CPU.

In this talk, we will explore training techniques and architecture modifications for training an efficient and compact network for person ReID, that achieves state-of-the-art results while having 10x fewer parameters and 13x fewer FLOPS.

Then, we will shortly present how we can apply our learning and techniques, such as the soft triplet loss, on other scenarios such as for the fine-grained image classification task.

Building fresh accurate maps of road items is a key ingredient in smart cities management and enabling fully autonomous vehicles. Building such maps from chip sensors such as monocular camera, GPS sensor and IMU, is a major challenge. It is even harder doing it in crowdsourcing setting, where the data is noisy and the camera position is arbitrary and unknown.

In this talk, we address this problem and related issues, namely; Camera alignment, self-localization, depth estimation, etc’. We demonstrate that using self-supervised approaches along with large corpus of diverse noisy-unlabeled data, we can get surprisingly accurate results.

We introduce a network that directly predicts the 3D layout of lanes in a road scene from a single image. This work marks a first attempt to address this task with on-board sensing without assuming a known constant lane width or relying on pre-mapped environments. Our network architecture, 3D-LaneNet, applies two new concepts: intra-network inverse-perspective mapping (IPM) and anchor-based lane representation. The intra-network IPM projection facilitates a dual-representation information flow in both regular image-view and top-view. An anchor-per-column output representation enables our end-to-end approach which replaces common heuristics such as clustering and outlier rejection, casting lane estimation as an object detection problem. In addition, our approach explicitly handles complex situations such as lane merges and splits. Results are shown on two new 3D lane datasets, a synthetic and a real one. For comparison with existing methods, we test our approach on the image-only tuSimple lane detection benchmark, achieving performance competitive with state-of-the-art. 

Given a pair of images of the same geographic area taken at different times, we wish to detect changes between them. Change detection is a challenging task. It is required to distinguish between fundamental changes, often man made, and insignificant natural ones. The latter may result from changing lighting, weather, camera pose, slight vegetation movement due to wind, and small errors in image registration. We address the change detection problem by training a learned descriptor using registered image pairs. Our fully convolutional CNN-based descriptor can efficiently detect changes in large aerial image pairs. It is shown to generalize well for a completely new scene and type of changes, while being robust to registration errors. The labeling of each image pair as similar or different is implied by the automatic registration process. Therefore, no manual annotation of any kind is required. While the lack of supervision results in label noise, the algorithm proves highly robust to it.

In today's increasingly visual world of e-commerce products are often accompanied by photo galleries describing various product aspects. We are going to deep dive into the travel accommodations use case and discuss a deep learning-based solution to the problem of finding meaningful representations of hotel galleries in a large scale e-commerce setting. The universality of embeddings and their flexibility to new downstream tasks is achieved through training the gallery encoder on multiple independent tasks using multi-task learning (MTL) approach. To evaluate the role of MTL in gallery encoding we look at how the performance of the joint MTL-trained model on each task compares to the model performances of separately trained end-to-end models. To assess the quality of learned representations we mainly look at their performance in downstream applications.

Monitoring large areas is presently feasible with high resolution drone cameras, as opposed to time-consuming and expensive ground surveys. In this work we reveal for the first time, the potential of using a state-of-the-art change detection GAN based algorithm with high resolution drone images for infrastructure inspection. We demonstrate this concept on solar panel installation. A deep learning, data-driven algorithm for identifying changes based on a change detection deep learning algorithm was proposed.

We use the Conditional Adversarial Network approach to present a framework for change detection in images. The proposed network architecture is based on pix2pix GAN framework. Extensive experimental results have shown that our proposed approach outperforms the other state-of-the-art change detection methods.

Computer Graphics images are commonly used in various fields like Medical imaging, gaming, animation, Augmented Reality and many more.

Contemporary Graphic Engines are able to produce scenes of limited photorealism.

In this talk, we will depict our method for generating visually appealing photorealistic images and temporally coherent videos from Computer generated scenes using Deep Learning.

In today's increasingly visual world of e-commerce products are often accompanied by photo galleries describing various product aspects. We are going to deep dive into the travel accommodations use case and discuss a deep learning-based solution to the problem of finding meaningful representations of hotel galleries in a large scale e-commerce setting. The universality of embeddings and their flexibility to new downstream tasks is achieved through training the gallery encoder on multiple independent tasks using multi-task learning (MTL) approach. To evaluate the role of MTL in gallery encoding we look at how the performance of the joint MTL-trained model on each task compares to the model performances of separately trained end-to-end models. To assess the quality of learned representations we mainly look at their performance in downstream applications.

Face alignment is a known computer vision challenge that has been explored enthusiastically in the past two decades. Thanks to the deep learning paradigm, researchers are now able to train networks to estimate 3D poses of objects using large amounts of data. A recurring limitation in these solutions is the lack of labelled data. Moreover, it is challenging to annotate 2D images with 3D labels. In this talk, we review the latest works on this topic and present how we generated massive amounts of semi-synthetic data using 3D morphable models. With this data, we were able to train a competitive real-time face tracker that is also lightweight and highly accurate for the task of head pose estimation.

Many deep learning models, developed in recent years, reach higher ImageNet accuracy than ResNet50, with fewer or comparable FLOPS count. While FLOPs are often seen as a proxy for network efficiency, when measuring actual GPU training and inference throughput, vanilla ResNet50 usually offers better throughput-accuracy trade-off.

In this talk, we will discuss the bottlenecks induced by FLOPs-optimizations, and suggest alternative design patterns that better utilize GPU structure and assets. We then introduce a new family of GPU-dedicated models, called TResNet, which achieves better accuracy and efficiency than previous ConvNets.

We demonstrate that on ImageNet, all along the top1 accuracy curve TResNet gives better GPU throughput than existing models. In addition, on three commonly used downstream single-label classification datasets it reaches new state-of-the-art accuracies. We also show that TResNet generalizes well to other computer vision tasks, reaching top scores on multi-label classification and object detection tasks.

In this talk I will introduce a novel Deep Learning framework, which quantitatively estimates image segmentation quality without the need for human inspection or labeling. We refer to this method as a Quality Assurance Network - QANet. Specifically, given an image and a ‘proposed’ corresponding segmentation, obtained by any method including manual annotation, the QANet solves a regression problem in order to estimate a predefined quality measure (or example the IoU or a Dice score) with respect to the unknown ground truth. The QANet is by no means yet another segmentation method. Instead, it performs a multi-level, multi-feature comparison of an image-segmentation pair based on a unique network architecture, called the RibCage.

To demonstrate the strength of the QANet, we addressed the evaluation of instance segmentation using two different datasets from different domains, namely, high throughput live cell microscopy images from the Cell Segmentation Benchmark and natural images of plants from the Leaf Segmentation Challenge. While synthesized segmentations were used to train the QANet, it was tested on segmentations obtained by publicly available methods that participated in the different challenges. We show that the QANet accurately estimates the scores of the evaluated segmentations with respect to the hidden ground truth, as published by the challenges’ organizers.

We introduce a new method of image-registration, named "semantic spatial alignment" (SSA).

This method performs an optimization of the semantic difference loss between two images, using a gradient-descend-process which optimizes the parameters of a neural-network composed of a single differentiable spatial-transformer. This new method shows a dramatic improvement over state-of-the-art feature-point-matching methods (e.g SIFT, ORB), when inputs are time-repeating orthomosaics of tree-plantations, where inputs can be from different sensors and resolutions, and contain changes in the shape of the tree objects. The method is also superior in cases where the success of affine, projective or other simple homographic transformation maps are limited. The method shows a successful use of deep learning in dramatically improving a traditional "classical computer vision task" as image-registration.

We introduce SinGAN, an unconditional generative model that can be learned from a single natural image. Our model is trained to capture the internal distribution of patches within the image, and is then able to generate high quality, diverse samples that carry the same visual content as the image. SinGAN contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. This allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image. In contrast to previous single image GAN schemes, our approach is not limited to texture images, and is not conditional (i.e. it generates samples from noise). User studies confirm that the generated samples are commonly confused to be real images. We illustrate the utility of SinGAN in a wide range of image manipulation tasks.

Face and head detection are fundamental problems in computer vision and related areas.
They are a common thread in contemporary applications such as augmented reality, biometric authentication, and social media.

The work presented in this lecture is a novel approach for head localization.

The suggested algorithm infers a person's head location given a prior face detection and other extracted features from the image.

Some features are mandatory and some are optional.

The algorithm uses classical methods – and thus it excels in computational efficiency – to solve the 2D to 2D problem via a 3D person polygon mesh, by understanding the most appropriate model and transformation.

We present a deep learning system for testing graphics units by detecting novel visual corruptions in videos. Unlike previous work in which manual tagging was required to collect labeled training data, our weak supervision method is fully automatic and needs no human labelling. This is achieved by reproducing driver bugs that increase the probability of generating corruptions, and by making use of ideas and methods from the Multiple Instance Learning (MIL) setting. In our experiments, we significantly outperform self-supervised methods such as GAN-based models and discover novel corruptions undetected by baselines, while adhering to strict requirements on accuracy and efficiency of our real-time system.

The chest radiography is by far the most commonly performed radiological examination for screening and diagnosis of many cardiac and pulmonary diseases. However, there is an immense world-wide decrease in the number of physicians capable of providing their rapid and accurate interpretation.  With 2 billion people joining the middle class worldwide and a growing global shortage of clinical experts, there is a sense of urgency to develop technologies which can help bridge the gap between supply and demand of radiology services.

Here we will review the research and development of Zebra-Medical AI based solutions aimed at providing automated and scalable diagnostic support in interpretation of chest radiographs.

We'll demonstrate the application of our technology on real life clinical examples where such solutions have impacted patients' care by substantially reducing time to treatment and preventing misdiagnosis.

As deep learning is showing potential value in different markets, there is an increasing need to be able to run inference efficiently on edge devices.

In this talk we will focus on the fundamental characteristics of deep learning algorithms, analyze the challenges they introduce to the classical 60 years old Von-Neuman processing approach and review the guidelines to building more efficient domain specific processing architecture.

Beginning with some theoretical reasoning behind domain-specific architectures and their implementation in the field of deep learning, and more specifically for machine vision applications. We will use various quantitative measures, and more detailed design examples in order to make a link between theory and practice.

Hailo has developed a specialized deep learning processor that delivers the performance of a data center-class computer to edge devices. Hailo’s AI microprocessor is the product of a rethinking of traditional computer architectures, enabling smart devices to perform sophisticated deep learning tasks such as imagery and sensory processing in real time with minimal power consumption, size and cost.