Track 1: Deep Learning Summit schedule

Lisha is a principal at Amplify Partners. She focuses on companies that leverage machine learning and data to solve problems and she is excited to be investing at a time when algorithmic and data-driven methods have such incredible potential for impact. Lisha completed her PhD at UC Berkeley focusing on deep learning and probability applied to the problem of clustering in graphs. Supported by the NSERC CGS fellowship, she worked with Prof. David Aldous and Prof. Joan Bruna. While at Berkeley she also did statistical consulting, advising on methods and analysis for experimentation and interpretation, and interned as a data scientist at Pinterest and Stitch Fix. She was the lecturer of discrete mathematics, as well as the graduate instructor for probability and statistics and intro CS theory.

Deep Learning for Computer Graphics: Learning to Estimate Lighting From Photographs

We propose an automatic method to infer high dynamic range illumination from a single, limited field-of-view, low dynamic range photograph of a scene. In contrast to previous work that relies on specialized image capture, user input, and/or simple scene models, we train an end-to-end deep neural network that directly regresses a limited field-of-view photo to HDR illumination, without strong assumptions on scene geometry, material properties, or lighting. This allows us to automatically recover high-quality HDR illumination estimates that significantly outperform previous state-of-the-art methods. Consequently, using our illumination estimates for applications like 3D object insertion, we can achieve compelling, photorealistic results.

Jean-François Lalonde is an Assistant Professor in Electrical and Computer Engineering at Laval University, Quebec City, since 2013. Previously, he was a Post-Doctoral Associate at Disney Research, Pittsburgh. He received a B.Eng. degree in Computer Engineering with honors from Laval University, Canada, in 2004. He earned his M.S at the Robotics Institute at Carnegie Mellon University in 2006 under Prof. Martial Hebert and received his Ph.D., also from Carnegie Mellon, in 2011 under the supervision of Profs. Alexei A. Efros and Srinivasa G. Narasimhan. His Ph.D. thesis won the 2010-11 CMU School of Computer Science Distinguished Dissertation Award. After graduation, he became a Computer Vision Scientist at Tandent, Inc., where he helped develop LightBrush™, the first commercial intrinsic imaging application. He also introduced intrinsic videos at SIGGRAPH 2012 while at Tandent. His research focuses on lighting-aware image understanding and synthesis by leveraging deep learning and large amounts of data.

Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift

Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive uncertainty. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.

Jasper Snoek completed his PhD in machine learning at the University of Toronto in 2013. He subsequently held postdoctoral fellowships at the University of Toronto, under Geoffrey Hinton and Ruslan Salakhutdinov, and at the Harvard Center for Research on Computation and Society, under Ryan Adams. Jasper co-founded the machine learning startup Whetlab, which was acquired by Twitter in 2015. Currently, he is a research scientist at Google Brain in Cambridge, MA.

Advances in Deep Architectures and Methods for Separating Vocals in Recorded Music

Source separation of audio mixtures, with an emphasis on the human voice, remains one of the enticing unsolved challenges in audio signal processing. This challenge is amplified in the context of recorded music, where often many sound sources are intentionally correlated in both time and frequency. In this talk, we present recent advances in the state of the art for separating singing voice and accompaniment in popular music audio recordings, leveraging semi-supervised datasets mined from a large commercial music catalog. In addition, we explore the effects of combining deep convolutional U-Net architectures with multi-task learning for vocal separation.

Eric J. Humphrey is a research scientist at Spotify, and acting Secretary on the board of the International Society for Music Information Retrieval (ISMIR). Previously, he has worked or consulted in a research capacity for various companies, notably THX and MuseAmi, and is a contributing organizer of a monthly Music Hackathon series in NYC. He earned his Ph.D. at New York University in Steinhardt's Music Technology Department under the direction of Juan Pablo Bello, Yann LeCun, and Panayotis Mavromatis, exploring the application of deep learning to the domains of audio signal processing and music informatics. When not trying to help machines understand music, you can find him running the streets of Brooklyn or hiding out in his music studio.

Few-Shot Learning: Thoughts On Where We Should Be Going

Few-shot learning is the problem of learning new tasks from little amounts of labeled data. This is achieved by performing a form of transfer learning, from the data of many other existing tasks. This topic has gained tremendous interest in the past few years, with several new methods being proposed each month. In this talk, I suggest we take a step back, look at what we have achieved and, most importantly, consider where this research should be going next.

Hugo Larochelle is Research Scientist at Google and Assistant Professor at the Université de Sherbrooke (UdeS). Before, he was working with Twitter and he also spent two years in the machine learning group at University of Toronto, as a postdoctoral fellow under the supervision of Geoffrey Hinton. He obtained his Ph.D. at Université de Montréal, under the supervision of Yoshua Bengio. He is the recipient of two Google Faculty Awards. His professional involvement includes associate editor for the IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), member of the editorial board of the Journal of Artificial Intelligence Research (JAIR) and program chair for the International Conference on Learning Representations (ICLR) of 2015 and 2016.

Deep Learning, Where Are You Going?

There are three axes along which advances in machine learning and deep learning happen. They are (1) network architectures, (2) learning algorithms and (3) spatio-temporal abstraction. In this talk, I will describe a set of research topics I've pursued in each of these axes. For network architectures, I will describe how recurrent neural networks, which were largely forgotten during 90s and early 2000s, have evolved over time and have finally become a de facto standard in machine translation. I continue on to discussing various learning paradigms, how they related to each other, and how they are combined in order to build a strong learning system. Along this line, I briefly discuss my latest research on designing a query-efficient imitation learning algorithm for autonomous driving. Lastly, I present my view on what it means to be a higher-level learning system. Under this view each and every end-to-end trainable neural network serves as a module, regardless of how they were trained, and interacts with each other in order to solve a higher-level task. I will describe my latest research on trainable decoding algorithm as a first step toward building such a framework.

Kyunghyun Cho is an assistant professor of computer science and data science at New York University. He was a postdoctoral fellow at University of Montreal until summer 2015, and received PhD and MSc degrees from Aalto University early 2014. He tries best to find a balance among machine learning, natural language processing and life, but often fails to do so.

Deep Learning in Speech Recognition

While neural networks had been used in speech recognition in the early 1990s, they did not outperform the traditional machine learning approaches until 2010, when Alex’s team members at Microsoft Research demonstrated the superiority of Deep Neural Networks (DNN) for large vocabulary speech recognition systems. The speech community rapidly adopted deep learning, followed by the image processing, and many other disciplines. In this talk I will explain the transition to deep learning, what the speech recognition field has accomplished, and the remaining challenges.

Alex Acero (PhD, Carnegie Mellon, 1990) is Sr. Director in the Siri team in charge of speech recognition, speech synthesis, and machine translation. Prior to joining Apple, he spent 20 years at Microsoft Research managing teams in speech, audio, multimedia, computer vision, natural language processing, machine translation, machine learning, and information retrieval. Dr. Acero is an IEEE Fellow and ISCA Fellow. Alex has served as President of the IEEE Signal Processing Society and is currently a member of the IEEE Board of Directors. He is the author of the textbook “Spoken Language Processing”. Dr. Acero has published over 250 technical papers and has over 150 US patents.

The Next Generation of AI Chips

Deep learning has fuelled significant progress in computer vision, speech recognition, and natural language processing. We have seen a single deep learning algorithm learn to recognize two vastly different languages, English and Mandarin, begin to synthesize realistic human speech, recognize visual data, and even to understand human language. At Baidu, we think that this is just the beginning, and high performance computing is poised to help. It turns out that deep learning is compute limited, even on the fastest machines that we can build. This talk will focus on a new discovery that significantly accelerates deep learning training by using mixed 16-bit and 32-bit IEEE standard floating point arithmetic. Unlike previous work in this area, the first generation of commodity hardware realizing an up to 8x speedup from this approach is already shipping in volume. We demonstrate the success of this approach across 15 state of the art deep learning training applications drawn from a diverse set of problem domains, and detail the small changes to deep learning frameworks needed to support this technology.

Greg Diamos leads computer systems research at Baidu’s Silicon Valley AI Lab (SVAIL), where he helped develop the Deep Speech and Deep Voice systems. Before Baidu, Greg contributed to the design of compiler and microarchitecture technologies used in the Volta GPU at NVIDIA. Greg holds a PhD from the Georgia Institute of Technology, where he led the development of the GPU-Ocelot dynamic compiler, which targeted CPUs and GPUs from the same program representation.

Towards Perceptual Machines That See, Converse, and Reason

A successful autonomous system needs to not only understand the visual world but also communicate its understanding with humans. To make this possible, language can serve as a natural link between high level semantic concepts and low level visual perception. In this talk, I'll discuss recent work in the domain of vision and language, covering topics such as image/video captioning and retrieval, and question-answering. I’ll also talk about our recent work on task execution via language instructions.

Sanja Fidler is an Assistant Professor at the Department of Computer Science, University of Toronto. Previously she was a Research Assistant Professor at TTI-Chicago, a philanthropically endowed academic institute located in the campus of the University of Chicago. She completed her PhD in computer science at the University of Ljubljana in 2010, and was a postdoctoral fellow at University of Toronto during 2011-2012. She has served in program committees of numerous international conferences, and has received three outstanding reviewer awards. Together with Rich Zemel and Raquel Urtasun, she received the NVIDIA Pioneer of AI award. Her main research interests are object detection, 3D scene understanding, and the intersection of language and vision.

Learning from Video: Recognizing Actions and Localizing Moments with Natural Language

This talk will describe two works in video understanding. In the first part, I will describe ActionVLAD, a new video representation for action classification that aggregates local convolutional features across an entire spatio-temporal extent of a video. In the second part, I will describe an approach that retrieves a specific temporal segment (moment) from a video given a natural language text description. We address lack of video datasets for this task by collecting the Distinct Describable Moments (DiDeMo) dataset which consists of over 10,000 unedited, personal videos in diverse visual settings with pairs of localized video segments and referring expressions.

Bryan Russell is currently a Research Scientist at Adobe Research in San Francisco, CA. He received his Ph.D. from MIT in the Computer Science and Artificial Intelligence Laboratory and was a post-doctoral fellow in the INRIA Willow team in Paris, France. He was a Research Scientist with Intel Labs as part of the Intel Science and Technology Center for Visual Computing (ISTC-VC) and was Affiliate Faculty at the University of Washington.

Unsupervised Domain Adaptation with Adversarial Network

This talk presents our recent work on using Adversarial learning to improve the recognition task in the presence of domain shifts or bias between the source (training) and target (testing) domains. Our new approach factors the learned feature space of both source domains and target domains into discriminative space and reconstructive space, where the discriminative space captures class-specific information while the reconstructive space captures the domain-specific information. A GAN-loss is used to minimize the domain shifts in the discriminative space between the two domains for better generalization performance. Our preliminary results show the promise of our approach, achieving new state-of-the-art results on standard cross-domain digit classification tasks.

Jianchao Yang is currently a Lead Research Scientist at Snap Inc. Before joining Snap, he was a Research Scientist at Adobe Research. He received his M.S. and Ph.D. degrees both from the ECE Department of University of Illinois at Urbana-Champaign, under supervision of Prof. Thomas Huang. His research focuses on computer vision, deep learning, and image and video processing. He has published more than 80 technical papers on top tier conferences and journals, with Google scholar citation more than 13,000 times. He received the Best Student Paper award from ICCV 2010, the winner prize of the classification task in PASCAL VOC 2009, first position for object localization using external data for ILSVRC ImageNet 2014, and third place in WebVision Challenge 2017. He serves as workshop chair for ACM MM 2017.

Composing Graphical Models With Neural Networks for Structured Representations and Fast Inference

How can we build structured, but flexible models? We propose a general modeling and inference framework that combines the complementary strengths of probabilistic graphical models and deep learning methods. Our model family combines latent graphical models with neural network observation models. All components are trained simultaneously with a single scalable stochastic variational objective. We illustrate this framework with several example models, and by showing how to automatically segment and categorize mouse behavior from raw video.

David Duvenaud is an assistant professor in computer science and statistics at the University of Toronto. His postdoc was at Harvard University, where he worked on hyperparameter optimization, variational inference, deep learning, and automatic chemical design. He did his Ph.D. at the University of Cambridge, studying Bayesian nonparametrics with Zoubin Ghahramani and Carl Rasmussen. David spent two summers in the machine vision team at Google Research, and also co-founded Invenia, an energy forecasting and trading company.

Divide and Conquer Networks

Many engineering and scientific tasks require solving algorithmic tasks at scale under computational constraints. In general, these constraints are hard to optimize in closed form, motivating the use of data-driven algorithmic learning. Rather than studying this as a black-box discrete regression problem with no assumption whatsoever on the input-output mapping, we concentrate on tasks that are amenable to the principle of divide and conquer, and study what are its implications in terms of learning. This principle creates a powerful inductive bias that can be leveraged with neural architectures that are defined recursively and dynamically, by learning two scale-invariant atomic operations: how to split a given input into smaller sets, and how to merge two partially solved tasks into a larger partial solution. Thanks to the dynamic aspect of such architecture, computational complexity can be incorporated as a regularization term that can be optimized by backpropagation. In this talk, we will illustrate the flexibility and efficiency of the Divide-and-Conquer Network on combinatorial and geometric tasks, such as sorting, clustering and convex hulls.

Joan Bruna graduated from Universitat Politecnica de Catalunya (Barcelona, Spain) in both Mathematics and Electrical Engineering. He obtained an M.Sc. in applied mathematics from ENS Cachan (France). He then became a research engineer in an image processing startup, developing real-time video processing algorithms. He obtained his PhD in Applied Mathematics at Ecole Polytechnique (France), under the supervision of Prof. Stephane Mallat. He was a postdoctoral researcher at the Courant Institute, NYU, and a postdoctoral fellow at Facebook AI Research. In 2015, he became Assistant Professor at UC Berkeley, Statistics Department, and starting Fall 2016 he joined the Courant Institute (NYU, New York) as Assistant Professor in Computer Science, Data Science and Mathematics (affiliated). His research interests include invariant signal representations, high-dimensional statistics and stochastic processes, deep learning and its applications to signal processing. He is co-chair of the IPAM Workshop on New Deep Learning Techniques (2018).

Join us for an evening of discussions & networking around the progress and application of machine intelligence from leading females working in the space following Day 2 of the Deep Learning Summit. The dinner is open to all and provides a unique opportunity to network with peers and make new contacts with fellow diners including Directors, CEOs, Data Scientists, Founders, Researchers and Engineers.

Enjoy a Champagne reception, followed by 3 courses of the finest cuisine and wine to match and compliment the evening and hear short talks from experts following each course.

To register for the event visit the website here or contact Katie on [email protected] for further enquiries.

Stepping Stones in an AI Strategy for the Workplace

In the past five years, startups and large companies have been building up their “artificial intelligence quotient” (AIQ). For big companies, AI presents the opportunity for business transformation. For entrepreneurs, it is the asymmetric tool they can use to take on much larger competitors. Join Accenture’s Global Head of AI for Products, Shyam Thyagaraj, as he shares some insights from Accenture’s most recent research and learn how to “boost your AIQ”. The theory will be followed by a real business use case. This session will include a casual discussion-style lecture from Shyam as well as a facilitated group ideation session to help participants define and share their own insights around how start-ups to big companies can better collaborate in this space to truly capture the value of AI.

Shyam Thyagaraj is a Managing Director in Accenture’s Consulting practice. Shyam leads Accenture’s Artificial Intelligence practice for Consumer and Industrial clients globally. Shyam and his team are focused on helping large enterprises innovate by mapping new technologies to industry specific business problems. His team helps enterprises develop the right strategy and enabling structures, including talent, data and infrastructure, required to achieve value from AI technologies.

Explainable by Design AI is the Real Challenge

Artificial Intelligence capabilities are set to become pervasive and to spread across a span a various systems including air, ground and sea transportation vehicles. Yet Machine Learning algorithms and their derived features do not provide any means to insure and enforce Safety and Security which is the number 1 driver and priority to certify new technologies on board critical transportation systems: “deployable AI can only be achieved with Explainable AI”. To cope with the inherent unpredictability and the lack of guarantees of Machine Learning black boxes, Industry, Academia and Regulations Agencies have to work jointly to design and develop frameworks and Explainable by design AI to allow deployable AI features onboard Aircrafts, Ships and any ground transportation vehicles.

David Sadek, holding a Doctorate in Computer Science, joins THALES as Vice President Research and Technology. He is a world renowned expert in Artificial Intelligence and Cognitive Science and Dr Sadek’s nominations to several past and current national and international steering and evaluation committees highlight his longstanding contributions to the sector. His impressive career includes a number of senior positions, notably with Orange and Orange Labs, as Vice President for Research & Strategy, as well as over a decade in Research & Development, designing intelligent agents and natural human-machine dialogue. His research led to the design and implementation of the first technologies in conversational agents, as well as the ACL inter-agent communication language standard.

Journey to the Cognitive Era with IBM - How to leverage Deep Learning to win the AI Arms Race

While Deep Learning has widely been accepted as one of the most exciting developments in computer science of the past fifty years, many organizations find themselves struggling with just how to get started. IBM's historical leadership in artificial intelligence dates back to the 1950s with AI pioneer Arthur Samuel's very first checkers playing machine learning system and extends to present day cutting edge AI research and ground breaking technology such as Watson, that have disrupted wide spread industries and many facets of society. This session will highlight some of IBM's recent accomplishments and will also illustrate that regardless of where you are in your journey to the cognitive era, IBM has the breadth of revolutionary technology offerings, and depth of data science expertise to help your organization leverage deep learning and win the AI arms race.

Adel El-Hallak is the Director of Product Management for Deep Learning within IBM’s Cognitive Systems unit. Adel’s mission is to democratize and operationalize deep learning so it can be more easily consumed in order to augment human intelligence. Adel’s previous roles include leadership positions across sales and marketing for high performance computing and advanced analytics. Adel holds a bachelor’s degree in computer science from McGill University in Canada and an M.B.A. from Warwick Business School in the U.K.

Journey to the Cognitive Era with IBM - How to leverage Deep Learning to win the AI Arms Race

While Deep Learning has widely been accepted as one of the most exciting developments in computer science of the past fifty years, many organizations find themselves struggling with just how to get started. IBM's historical leadership in artificial intelligence dates back to the 1950s with AI pioneer Arthur Samuel's very first checkers playing machine learning system and extends to present day cutting edge AI research and ground breaking technology such as Watson, that have disrupted wide spread industries and many facets of society. This session will highlight some of IBM's recent accomplishments and will also illustrate that regardless of where you are in your journey to the cognitive era, IBM has the breadth of revolutionary technology offerings, and depth of data science expertise to help your organization leverage deep learning and win the AI arms race.

Dr. Gschwind leads the development of IBM's cognitive enterprise solution to deliver improved business results with advanced data understanding and AI-based modeling in the IBM Global Chief Data Office. He leads the design and implementation of IBM's Deep Learning at Scale Platform coupled to IBM's Corporate Enterprise Data Platform, and its applications. As Chief engineer for Machine and Deep Learning in IBM's Systems Group, he initiated the development of PowerAI and AI solutions in based on a Deep Learning optimized hardware and software systems. He previous servers as a chief architect for supercomputers, enterprise servers and game console processors.