Deep Learning Summit schedule

I joined Playfair Capital in 2013 to focus on deal sourcing, due diligence, and helping our companies grow. I'm particularly interested in artificial intelligence and machine learning, infrastructure-as-a-service, mobile, and bioinformatics. I've led, originated or participated in Seed through Growth investments including Mapillary, Appear Here, Dojo, and Festicket.

Prior to Playfair, I earned an M.Phil and Ph.D in oncology as a Gates and Dr. Herchel Smith Scholar at the University of Cambridge, and a BA in biology from Williams College. I've published research focused on technologies to halt the fatal spread of cancer around the body.

Bringing Deep Learning to The Front Lines of Healthcare

Today’s healthcare system is not built for the seamless integration of innovations in machine learning. Health data is heavily siloed, inaccessible, and non standardized, making it challenging to work with in machine learning systems. Additionally, deep learning techniques aren’t well suited for many healthcare problems due to the difficulty in interpreting their decisions. At Remedy, we’re building a healthcare system that captures granular data at the point of care, and enables the deployment of machine learning models at the point of care. We’re also developing interpretable model to tackle tasks such as diagnosis, physician education, treatment planning, and triage.

Will is CEO of Remedy Health, a startup focused on building a healthcare system around a strong software backbone to enable the seamless integration of new innovations into care. Will is also a venture partner at Alsop Louie Partners. Prior to this Will worked in R&D on SpaceX’s internet satellite project, and studied physics and computer science at MIT. An Ohio native, Will spent his childhood developing a particle collider in his basement, using the nuclear reactions it carried out to investigate novel methods of medical imaging.

Thanks to deep learning, AI algorithms can now surpass human performance in image classification. However, behind these results lie tens of thousands of man hours spent annotating images. This significantly prohibits commercial applications where cost and time to market are key. At Tractable, our solution centers on creating a feedback loop from learning algorithm to human, turning the latter into a “teacher” rather than a blind labeler. Dimensionality reduction, information retrieval and transfer learning are some of our core proprietary techniques. We will demonstrate a 15x labeling cost reduction on the expert task of estimating from images the cost to repair a damaged vehicle – an important application for the insurance industry.

Alex is Co-founder & CEO of Tractable, a young London-based startup bringing recent breakthroughs in AI to industry. Tractable's current focus is on automating visual recognition tasks. Its long term vision is to expand into natural language, robot control, and spread disruptive AI throughout industry. Tractable was founded in 2014 and is backed by $2M of venture capital from Silicon Valley investors, led by Zetta Venture Partners. Alex has a degree in econometrics & mathematical economics from the LSE, and a postgraduate degree in computer science from Imperial College London. Alex's experience within Deep Learning investing is on the receiving side, particularly on how to attract US venture capital into Europe as early as the seed stage.

Is solving video the key next breakthrough in computer vision? We’ll discuss the key challenges in applying deep learning techniques to video understanding. This will include approaches to building high quality datasets and annotating data for video is quite different than image understanding. What are the key use cases for video today and tomorrow? How do we address concerns around privacy and fears about “big brother”? Last but not least how does video advance the field of AI more towards general intelligence and common sense understanding of the physical world in machine learning models..

Roland Memisevic received his PhD in Computer Science from the University of Toronto in 2008. He subsequently held positions as research scientist at PNYLab, Princeton, as post-doctoral fellow at the University of Toronto and ETH Zurich, and as junior professor at the University of Frankfurt. In 2012 he joined the MILA deep learning group at the University of Montreal as assistant professor. He has been on leave from his academic position since 2016 to lead the research efforts at Twenty Billion Neurons, a German-Canadian AI startup he co-founded. Roland is Fellow of the Canadian Institute for Advanced Research (CIFAR).

Why Do Corporations Need to Own Their Data?

Augustin Marty is the CEO of Deepomatic, a company that believes artificial intelligence should be made accessible to all. To achieve this goal, Deepomatic has created a platform which enables businesses to develop their own image recognition systems. By 28 he had founded his first company in China, had worked in India, optimising combustion cycles for Power Plants, and at Vinci Construction Group designing and selling engineering projects. Augustin met his cofounders just after high school; sharing the same passion for entrepreneurship they decided to partner in early 2014 and created Deepomatic, the image intelligence company.

Systems Challenges for Deep Learning

Training neural network models and deploying them in production poses a unique set of computing challenges. The ability to train large models fast, allows researchers to explore the model landscape quickly and push the boundaries of what is possible with Deep Learning. However a single training run often consumes several exaflops of compute and can take a month or more to finish. Similarly, some problem areas like speech synthesis and recognition have real time requirements which places a limit on how much time it can take to evaluate a model in production. In this presentation, I will talk about three systems challenges that need to be addressed so that we can continue to train and deploy rich neural network models.

Shubho is now working on AI Research at FAIR. He was previous a senior research scientist at Silicon Valley AI Lab (SVAIL) at Baidu Research.

I am an architect of the High Performance Computing inspired training platform that is used to train some of the largest recurrent neural network models in the world at SVAIL. I also spend a large part of my time exploring models for doing both speech recognition and speech synthesis and what it would take to train these model at scale and deploy them to hundreds of millions of our users. I am the primary author of the WarpCTC project that is used commonly for speech recognition. Before coming to the industry, I got my PhD in Computer Science from UCDavis focusing on parallel algorithms for GPU computing and subsequently went to Stanford for a Masters in Financial Math.

Generalization and the Role of Data in Reinforcement Learning

Over the past decade, we have witnessed a revolution in supervised machine learning, as large, high-capacity models trained on huge datasets attain amazing results across a range of domains, from computer vision to natural language processing and speech recognition. But can these gains in supervised learning performance translate into more effective and optimal decision making? The branch of machine learning research that studies decision making is called reinforcement learning, and while more effective and performant reinforcement learning methods have also been developed over the past decade, in general it has proven challenging for reinforcement learning to benefit from large datasets, because it is conventionally thought of as an active online learning framework, which makes reusing large previously collected datasets difficult. In this talk, I will discuss how reinforcement learning algorithms can enable broad generalization through the use of large and diverse prior datasets. This concept lies at the core of offline reinforcement learning, which addresses the development of reinforcement learning methods that do not require active interaction with their environment but instead, much like current supervised learning methods, learn from previously collected datasets. Crucially, unlike supervised learning, such methods directly optimize for optimal downstream decision making, maximizing long-horizon reward signals. I will describe the computational and statistical challenges associated with offline reinforcement learning, describe recent algorithmic developments, and present a few promising applications.

Sergey Levine received a BS and MS in Computer Science from Stanford University in 2009, and a Ph.D. in Computer Science from Stanford University in 2014. He joined the faculty of the Department of Electrical Engineering and Computer Sciences at UC Berkeley in fall 2016. His work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning algorithms. Applications of his work include autonomous robots and vehicles, as well as computer vision and graphics. His research includes developing algorithms for end-to-end training of deep neural network policies that combine perception and control, scalable algorithms for inverse reinforcement learning, deep reinforcement learning algorithms, and more.

Machine Learning for Sustainability

Policies for sustainable development entail complex decisions about balancing environmental, economic, and societal needs. Making such decisions in an informed way presents significant computational challenges. Modern AI techniques combined with new data streams have the potential to yield accurate, inexpensive, and highly scalable models to inform research and policy. In this talk, I will present an overview of my group's research on applying computer science techniques in sustainability domains, including poverty and food security.

Stefano Ermon is an Assistant Professor in the Department of Computer Science at Stanford University, where he is affiliated with the Artificial Intelligence Laboratory and the Woods Institute for the Environment. Stefano's research is centered on techniques for scalable and accurate inference in graphical models, statistical modeling of data, large-scale combinatorial optimization, and robust decision making under uncertainty, and is motivated by a range of applications, in particular ones in the emerging field of computational sustainability.

More than a GPU: Platforms for Deep Learning

Training and deploying state of the art deep neural networks is very computationally intensive, with tens of exaflops needed to train a single model on a large dataset. The high density compute afforded by modern GPUs has been key to many of the advances in AI over the past few years. However, researchers need more than a fast processor – they also need optimized libraries, and tools to efficiently program so that they can experiment with new ideas. They also need scalable systems that use many of these processors together to train a single model. In this talk, I’ll discuss platforms for deep learning, and how NVIDIA is working to make the platforms of the future for Deep Learning.

Bryan Catanzaro is VP of Applied Deep Learning Research at NVIDIA, where he leads a team solving problems in fields ranging from video games to chip design using deep learning. Prior to his current role at NVIDIA, he worked at Baidu to create next generation systems for training and deploying end-to-end deep learning based speech recognition. Before that, he was a researcher at NVIDIA, where he wrote the research prototype and drove the creation of CUDNN, the low-level library now used by most AI researchers and deep learning frameworks to train neural networks. Bryan earned his PhD from Berkeley, where he built the Copperhead language and compiler, which allows Python programmers to use nested data parallel abstractions efficiently. He earned his MS and BS from Brigham Young University, where he worked on computer arithmetic for FPGAs.

A General Framework for Domain Adversarial Learning

Deep convolutional networks have provided significant advances in recent years, but progress has primarily been limited to fully supervised settings, requiring large amounts of human annotated data for training. Recent results in adversarial adaptive representation learning demonstrate that such methods can also excel when learning in sparse/weakly labeled settings across modalities and domains. In this talk, I will present a general framework for domain adversarial learning, by which a game is played between a discriminator which seeks to determine whether an image arises from the large labeled data source or from the new sparsely labeled dataset, while the representation seeks to limit the distinguishability of the two data sources. Together this game produces a model adapted for the new domain with minimal or no new human annotations.

Judy Hoffman is a Postdoctoral Researcher in the Stanford Computer Vision group, working with Fei-Fei Li. Her research focuses on developing learning representations and recognition models with limited human annotations. She received her PhD in Electrical Engineering and Computer Science from University of California, Berkeley in Summer 2016, where she was advised by Trevor Darrell and Kate Saenko. She is interested in lifelong learning, adaptive methods, and adversarial models.

Practical, Active, Interpretable & Deep Learning

I'll review applied deep learning techniques we use at Stitch Fix to understand our client's personal style. Interpretable deep learning models are not only useful to scientists, but lead to better client experiences -- no one wants to interact with a black box virtual assistant. We do this in several ways. We've extended factorization machines with variational techniques, which allows us to learn quickly by finding the most polarizing examples. And by enforcing sparsity in our models, we have RNNs and CNNs that reveal how they function. The result is a dynamic machine that learns quickly and challenges our client's style.

Chris Moody came from a Physics background from Caltech and UCSC, and is now a scientist at Stitch Fix's Data Labs. He has an avid interest in NLP, has dabbled in deep learning, variational methods, and Gaussian Processes. He's contributed to the Chainer deep learning library (http://chainer.org/), the super-fast Barnes-Hut version of t-SNE to scikit-learn (http://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html) and written (one of the few!) sparse tensor factorization libraries in Python (https://github.com/stitchfix/ntflib). Lately he's been working on lda2vec (https://lda2vec.readthedocs.org/en/latest/).

Quantifying Visual Aesthetics on Flickr

What makes an image beautiful? More pragmatically, can an algorithm distinguish high-quality photographs from casual snapshots to improve search results, recommendations, and user engagement at scale? We leverage social interaction with Flickr photos to generate a massive dataset for computational aesthetics and train machine learning models to predict the likelihood images being of high, medium, or low quality. I will present our approach and findings, address some of the challenges of quantifying subjective preferences, and discuss applications of the aesthetics model to finding, sharing, and creating visually compelling content in an online community.

Stacey Svetlichnaya is a software engineer on the Yahoo Vision & Machine Learning team. Her recent deep learning research includes object recognition, image aesthetic quality and style classification, photo caption generation, and modeling emoji usage. She has worked extensively on Flickr image search and data pipelines, as well as automating content discovery and recommendation. Prior to Flickr, she helped develop a visual similarity search engine with LookFlow, which Yahoo acquired in 2013. Stacey holds a BS and MS in Symbolic Systems from Stanford University.

Personalized Content/Image Selection

A decade ago, Netflix launched a challenge to predict how each user would rate each movie in our catalog. This accelerated the science of machine learning and matrix factorization. Since then, our learning algorithms and models have evolved with multiple layers, multiple stages and nonlinearities. Today, we use machine learning and deep variants to rank a large catalog by determining the relevance of each of our titles to each of our users, i.e. personalized content selection. We also use machine learning to find how to best present the top ranked items for the user. This includes selecting the best images to display for each title just for you, i.e. personalized image selection.

Tony directs machine learning research at Netflix and is sabbatical professor at Columbia University. He serves as general chair of the 2017 International Conference on Machine Learning. He has published over 100 scientific articles in the field of machine learning and has received several best paper awards.

Learning from Multi-Agent, Emergent Behaviors in a Simulated Environment

A revolution in reinforcement learning is happening, one that helps companies harness the more diverse, complex, virtual simulations available to accelerate the pace of innovation. Join this session to learn about particular environments already created that have yielded surprising advances in AI agents, and to better understand how emergent behaviors and open-endedness in multi-agent systems can lead to the most optimal designs and real-world practices.

Dr. Danny Lange is Senior Vice President of Artificial Intelligence and Machine Learning at Unity. As head of machine learning at Unity, Lange leads the company’s innovation around AI (Artificial Intelligence) and Machine Learning, focusing on bringing AI to simulation and gaming.

Prior to joining Unity, Lange was the head of machine learning at Uber, where he led efforts to build the world’s most versatile Machine Learning platform to support the company’s hyper-growth. Lange also served as General Manager of Amazon Machine Learning -- an AWS product that offers Machine Learning as a Cloud Service. Before that, he was Principal Development Manager at Microsoft where he led a product team focused on large-scale Machine Learning for Big Data.

Lange spent 8 years on Speech Recognition Systems, first as CTO of General Magic, Inc., then through his work on General Motor’s OnStar Virtual Advisor, one of the largest deployments of an intelligent personal assistant until Siri. Danny started his career as a Computer Scientist at IBM Research.

He holds MS and Ph.D. degrees in Computer Science from the Technical University of Denmark. He is a member of the Association for Computer Machinery (ACM) and IEEE Computer Society, and has several patents to his credit.

Shohei Hido received M.S in Informatics from Kyoto University in Japan, 2006. Since then, he has worked at IBM Research in Tokyo for six years as a staff researcher in machine learning and its applications to industries. After joining Preferred Infrastructure, Inc. in 2012, he has worked as the leader of Jubatus project, an open source software framework for real-time, streaming machine learning. After being assigned as Chief Research Officer of Preferred Networks, the spin-off company, he is currently responsible for Deep Intelligence in-Motion, software platform for using deep learning in IoT applications.