Machine Intelligence Summit schedule

Alex Brokaw is a writer and consultant based in New York City. He spent many years as a finance and technology reporter, most recently at The Verge. He now specializes in writing and creative services for artificial intelligence and robotics companies. Alex is passionate about helping these companies explain their technologies to customers and larger audiences, whether that’s through technical writing, marketing copy, or overall brand and marketing strategy.

End-to-end Machine Intelligence

The common perception of machine learning is that it starts with data and ends with a model. In real-world production systems, the traditional data science and machine learning workflow of data preparation, feature engineering and model selection, while important, is only one aspect. True machine intelligence requires continuous, real-time learning by closing the feedback loop. Equally important for enterprise-grade and mission-critical systems are the typical elements of software systems, such as scalability, fault tolerance and security.

Nick is a Principal Engineer at IBM’s Spark Technology Center. He is a member of the Apache Spark PMC and author of Machine Learning with Spark. Previously, he co-founded Graphflow, a startup focused on recommendations and customer intelligence. He has worked at Goldman Sachs, Cognitive Match, and led the Data Science team at Mxit, Africa’s largest social network. He is passionate about combining commercial focus with machine learning and cutting-edge technology to build intelligent systems that learn from data to add business value.

On the Path to Non-Biological Sentience

At Kindred our mission is to create human-like intelligence in machines, and our core hypothesis is that intelligence requires a body. We build robotic bodies of many different form factors and imbue them with new types of AI minds inspired by biological creatures’ brains. We pre-train these algorithms by demonstration, using large amounts of data from human operators, and craft reward functions such that the entities can subsequently become fully autonomous agents. Breakthroughs in robotics, increases in computational power, and advancements in machine learning algorithms are finally enabling us to train AI minds to have a more “general” form of intelligence. This presentation will describe our technology and address some of the questions that might arise as we begin to share our world with increasingly smarter robots.

Suzanne is co-founder and CSO of Kindred. She oversees the design and engineering of the company's human-like robots and is responsible for the development of cognitive architectures that allow these robots to learn about themselves and their environments. Before founding Kindred, Suzanne worked as a physicist at D-Wave, designing and building superconducting quantum processors, and as a researcher in quantum artificial intelligence software applications. Suzanne likes science outreach, retro tech art, coffee, cats, electronic music and extreme lifelogging. She is a published author of a book of art and poetry. She is passionate about robots and their role as a new form of symbiotic life in our society. Suzanne received her Ph.D. in experimental physics from the University of Birmingham (UK) in 2008, specializing in quantum device physics, microfabrication techniques, and low-temperature measurements of novel superconducting circuits.

Search Ranking And Personalization at Airbnb

Search ranking is a fundamental problem of crucial interest to major Internet companies, including web search engines, content publishing websites and marketplaces. However, despite sharing some common characteristics a one-size-fits-all solution does not exist in this space. Given a large difference in content that needs to be ranked and the parties affected by ranking, each search ranking problem is somewhat specific. Correspondingly, search ranking at Airbnb is quite unique, being a two-sided marketplace in which one needs to optimize for host and guest preferences, in a world where a user rarely consumes the same item twice and one listing can accept only one guest for a certain set of dates. In this talk, I will discuss challenges we have encountered and Machine Learning solutions we have developed for listing ranking at Airbnb. Specifically, the listing ranking problem boils down to prioritizing listings that are appealing to the guest but at the same time demoting listings tha t would likely reject the guest, which is not easily solvable using basic matrix completion or a straightforward linear model. I will shed the light on how we jointly optimize the two objectives by leveraging listing quality, location relevance, reviews, host response time as well as guest and host preferences and past booking history. Finally, we will talk about our recent work on using neural network models to train listing, query and user embeddings for purposes of enhancing search relevance and personalization, two core concepts in any modern search.

Mihajlo Grbovic, Ph.D. is a Senior Machine Learning Scientist on the Search Ranking Team at Airbnb. Prior to that, he was a Senior Research Manager at Yahoo Labs working on Advertising Sciences. He has more than 10 years of technical experience in applied Machine Learning, acting as a Science Lead in a portfolio of advertising technology projects on Monetization of Tumblr, Yahoo Email and Yahoo Search. Some of his biggest accomplishments include building a large scale Interest and Gender Targeting Pipeline for Tumblr, training Email Classifiers used in Yahoo Mail Smart Views that millions of people interact with every day, and introducing the next generation query-ad matching algorithm to Yahoo Sponsored Search. Dr. Grbovic published more than 40 peer-reviewed publications at top Machine Learning and Web Science Conferences and co-authored more than 10 pending patents. His work was featured in Wall Street Journal, Scientific American, MIT Technology Review, Popular Science and Market Watch.

Deep Learning for Climate Science

Climate change poses a major challenge for humanity in the 21st century. Characterizing the impact of climate change requires sophisticated analysis of complex datasets produced by climate simulations. In recent years, we have shown that Deep Learning based techniques are capable of classifying, detecting and segmenting extreme weather patterns in O(10) TB sized datasets. We have successfully scaled Deep Learning architectures to the largest CPU- and GPU-based HPC systems in the world. This talk will review our latest results, and conclude with efforts to create ClimateNet - an open labeled dataset and reference architecture for collaboration with the broader community.

Key Takeaways:

• Deep Learning can be applied to climate science problems
• Analyzing large datasets requires HPC systems
• Access to labeled data is a key challenge

Prabhat leads the Data and Analytics Services team at NERSC; his group is responsible for supporting over 7000 scientific users on NERSC’s HPC systems. His current research interests include Deep Learning, Machine Learning, Applied Statistics and High Performance Computing. Prabhat has co-authored over 150 papers spanning several domain sciences and topics in computer science. He has won 5 Best Paper Awards, 3 Industry Innovation Awards, and he was a part of the team that won the 2018 Gordon Bell Prize for their work on ‘Exascale Deep Learning’.

Beating Perils of Non-convexity: Guaranteed Training of Neural Networks using Tensor Methods

Neural networks have revolutionized performance across multiple domains such as computer vision and speech recognition. However, training a neural network is a highly non-convex problem and the conventional stochastic gradient descent can get stuck in spurious local optima. We propose a computationally efficient method for training neural networks that also has guaranteed risk bounds. It is based on tensor decomposition which is guaranteed to converge to the globally optimal solution under mild conditions. We explain how this framework can be leveraged to train feedforward and recurrent neural networks.

Hanie Sedghi is a Research Scientist at Allen Institute for Artificial Intelligence (AI2). Her research interests include large-scale machine learning, high-dimensional statistics and probabilistic models. More recently, she has been working on inference and learning in latent variable models. She has received her Ph.D. from University of Southern California with a minor in Mathematics in 2015. She was also a visiting researcher at University of California, Irvine working with professor Anandkumar during her Ph.D. She received her B.Sc. and M.Sc. degree from Sharif University of Technology, Tehran, Iran.

Who Said What: Modeling individual Labelers Improves Classification

Data are often labeled by many different experts, with each expert labeling a small fraction of the data and each sample receiving multiple labels. When experts disagree, the standard approaches are to treat the majority opinion as the truth or to model the truth as a distribution, but these do not make any use of potentially valuable information about which expert produced which label. We propose modeling the experts individually and then learning averaging weights for combining them, possibly in sample-specific ways. This allows us to give more weight to more reliable experts and take advantage of the unique strengths of individual experts at classifying certain types of data. We show that our approach performs better than three competing methods in computer-aided diagnosis of diabetic retinopathy.

Melody is a deep learning resident at Google Brain. Previously she interned as a trader at D. E. Shaw and conducted stem cell research in Doug Melton's lab. She received an M.A. in Statistics and B.A. in Chemistry and Physics from Harvard University, where she graduated with highest ranking. In her youth she medalled at the International Physics, Chemistry, and Biology Olympiads and was invited to the Canadian International Math Olympiad training camp. Melody has published in The Huffington Post, Harvard Political Review, and The Harvard Crimson. She is constantly fawning over music, reading rationality and psychoanalysis blogs, and enjoying long random walks along undirected paths and other outdoor activities.

An Application of Gradient Boosted Decision Trees (GBDT) for Query Runtime Prediction

In 2016, Uber's went through an exponential growth. To support such an unprecedented growth in business, Uber had to rapidly scale out its infrastructure. The rapid growth in the core business and the number of products offered by Uber coupled with an increase in the number of data scientists and analysts meant an increasing load on our data systems. Specifically, having a stable and an efficient data warehouse was a fundamental requirement to sustain the pace at which Uber was growing and operating in the marketplace. The starting point of every data-driven decision at Uber typically begins with fetching data using our internal web-based tool called Querybuilder. More than 150K SQL queries are issued every week on our data warehouse through Querybuilder. To manage this load efficiently, it is very important to be able to accurately predict the runtime of the query before it hits the warehouse. In this study, we share our approach on how we used a Gradient Boosted Decision Tree model for predicting the runtime class of a query (short query or long query). Further, the predicted label was used to route the incoming query into an appropriate queue for execution in real-time. This approach of routing the queries based on the predictions from the classifier increased the overall efficiency of our warehouse. We observed a significant decrease in the average waiting for a query execution and increased the throughput of our system under peak load.

Abi Komma is a Senior Data Scientist in the Applied Machine Learning group at Uber. He focuses on applying regression, classification, and survival analysis techniques to build predictive models for various business problems at Uber. He holds a Master's degree in Transportation Engineering from University of Florida and a Bachelor’s degree in Civil Engineering from Indian Institute of Technology (IIT) Madras.

On-device Neural Networks for Natural Language Processing

Deep neural networks reach state-of-the-art performance for wide range of Natural Language Processing, Computer Vision and Speech applications. Yet, one of the biggest challenges is running these complex networks on devices with tiny memory footprint and low computational capacity such as mobile phones, smart watches and Internet of Things. In this talk, I will introduce novel on-device Self-Governing Neural Networks (SGNNs), which learn compact projection vectors with local sensitive hashing. The key advantage of SGNNs over existing work is that they surmount the need for pre-trained word embeddings and complex networks with huge parameters. I will showcase results from extensive evaluations on wide range of natural language tasks such as dialog act classification and user intent prediction. Our findings show that SGNNs are effective at capturing low-dimensional semantic text representations, while maintaining high accuracy and improving over state-of-the-art results.

Dr. Zornitsa Kozareva is a Manager at Google, leading and managing the Natural Language Understanding group and efforts in Google Apps Intelligence. Prior to that, Dr. Kozareva was Manager of Amazon’s AWS Deep Learning group that built and launched the Natural Language Processing and Dialog services Amazon Comprehend and Amazon Lex. Dr. Kozareva was as Senior Manager at Yahoo! leading the Query Processing group that powered Mobile Search and Advertisement. From 2009 to 2014, Dr. Kozareva wore an academic hat as Research Professor at the University of Southern California CS Department with affiliation to Information Sciences Institute, where she spearheaded research funded by DARPA and IARPA on learning to read, interpreting metaphors and building knowledge bases from the Web. Dr. Kozareva regularly serves as Area Chair and PC of top tier NLP and ML conferences. Dr. Kozareva has organized four SemEval scientific challenges and has published over 80 research papers. Dr. Kozareva is a recipient of the John Atanasoff Award given by the President of Republic of Bulgaria in 2016 for her contributions and impact in science, education and industry; the Yahoo! Labs Excellence Award in 2014 and the RANLP Young Scientist Award in 2011.

Balancing Discovery & Continuation in Recommendations

When a member logs on to Netflix, she may be in one or a combination of different watching modes: discovering a new content, continuing to watch a partially-watched movie or a TV show, playing one of the contents she had put in her play list during an earlier session, etc. If, for example, we can reasonably predict when a member is more likely to be in the continuation mode, and which videos she is more likely to resume, it makes sense to place those videos in more prominent places of the home page. In this talk we focus on understanding the discovery vs. continuation behavior and explain how we have used machine learning to improve the member experience by learning a personalized balance between those two modes.

Hossein Taghavi received the B.Sc. degree from Sharif University of Technology, Tehran, Iran, in 2003, and the Ph.D. degree from the University of California at San Diego, La Jolla, in 2008, in electrical engineering. From 2008 to 2011, he was with Qualcomm, Inc. in San Diego, CA, conducting research and development on algorithms and systems for wireless networks. From 2011 to 2013, he was with Opera Solutions, LLC., in San Diego, CA, applying machine learning and data analytics to a variety of business applications. Since 2013, he has been with Netflix, Inc., contributing to algorithms and software driving the personalization and recommendation systems for Netflix members. His current interests include machine learning, recommendation systems, and distributed computing.

Adversarial Learning for Fine-Grained Image Search

While computer vision has been extensively studied, it still remains a challenging problem. In particular, fine-grained image search is a challenging problem due to the difficulty in capturing subtle differences regardless of pose variations of objects from fine-grained categories. Numerous algorithms using deep neural networks have achieved state-of-the-art

performance on fine-grained categorization, but they are not directly applicable to fine-grained image search. In this presentation, eBay’s Chief Scientist for Computer Vision, Robinson Piramuthu will present on eBay research that proposes a solution called FGGAN, which learns discriminative representations by implicitly studying geometric transformation from multi-view images for fine-grained image search.

As Chief Scientist for Computer Vision, Robinson drives eBay’s computer vision science strategy. With over 20 years of experience in computer vision, his expertise includes large scale visual search, coarse and fine-grained visual recognition, object detection, computer vision for fashion, 3D cues from 2D images, figure-ground segmentation and deep learning for vision, among other topics. Before joining eBay in 2011, he received his PhD in Electrical Engineering and Computer Science from the University of Michigan in 2000 specializing in information theory and statistical image processing. He also has a MS in control theory from the University of Florida, specializing in robust and nonlinear control systems.

AlphaStar: Mastering the Real-Time Strategy Game StarCraft II

Deep reinforcement learning approaches have been shown to perform well on domains where tasks and rewards and well-defined. However, in adversarial multi-agent environments, where the agent is required to improve its policy through self-play, the agent should not only solve the given task (i.e., learning to beat itself via self-play) but also develop diverse policies and strategies over time in order to become strong and robust when playing against unseen competitors. In this talk, I will present AlphaStar which is the first AI to defeat a top professional player in the game of Starcraft, one of the most challenging Real-Time Strategy (RTS) games. Specifically, I will show how such complex and robust strategies can emerge through a distributed multi-agent RL algorithm, where a population of agents compete with each other with slightly different internal goals.

Key Takeaways:

• The current state-of-the-art RL algorithms can achieve super-human performance on a complex real-time strategy game (Starcraft)

Junhyuk Oh is a research scientist at DeepMind. He received his Ph.D. from Computer Science and Engineering at the University of Michigan in 2018, co-advised by Prof. Honglak Lee and Prof. Satinder Singh. His research focuses on deep reinforcement learning problems such as dealing with partial observability, generalization, planning, and multi-agent reinforcement learning. His work was featured at MIT Technology Review and Daily Mail.

Nick is a Senior Data Scientist at CrowdFlower, where he works primarily to help build their new machine learning offering, CrowdFlower AI. Prior to CrowdFlower, he worked as a data scientist at SF text analytics startup Idibon. He has a Ph.D. from the University of Texas at Austin, where his research focused on human language comprehension and the construction of datasets for NLP applications. In his spare time he fixes bikes and collaborates on work applying cognitive science principles to the public health domain.

Kevin is responsible for the global capabilities of the AirMap product portfolio. His responsibilities include defining strategic and tactical capabilities roadmaps, managing the product team, and coordinating with the various development teams. Kevin joined AirMap after 15 years with Lockheed Martin working on Air Traffic Control systems as a Systems Architect and working as the Aviation Chief Technologist for all global aviation projects.

As the founding engineer of Reddit, Chris Slowe helped build it, sell it, and has recently come back with the other co-founders to fix it. Though a software engineer by vocation, his first attempt at a career started with his finishing a PhD in experimental physics where he learned about the importance of modeling, critical thinking, statistics, and (honestly) welding. He uses the first three skills regularly. Chris currently acts as Director of Foundation Engineering at Reddit, Inc.

This discussion will center around best practices, key challenges and “war stories” for turning data science and machine learning projects into production systems. Some key questions that we'll discuss include:

How to go from exploratory data science and “proof of concept” projects to production systems with all their requirements of scalability, fault tolerance, performance monitoring & feedback?

How to bridge the gaps between the teams involved in each aspect of the Data Product Continuum, and the differing requirements of each phase?

What are some of the best practices in productionizing data science and machine learning workflows?

What role do standards play (e.g. data interchange formats, model evaluation standards such as PMML and PFA)?

What is the state of systems catering to this need, whether open-source or proprietary?

What are the important tools and software systems?

What are the important processes and non-software systems?

What are practitioners currently using in their production systems? How have they solved these problems and what are the “pro tips” and “lessons learnt” that can be shared?

Music Discovery at Pandora

Finding the music of the moment can often be a challenging problem, even for humans with well-versed musical tastes. These challenges further explode into a myriad of complexities when attempting to construct algorithmic approaches for automatic playlist generation. A variety of factors play a role in influencing a listener’s perception of what music is appropriate on a given seed (e.g., musicological, social, geographical, generational), and these factors vary across different contexts and listeners. The talk will present an overview of recommendation at Pandora, followed by a deep dive into the challenges of recommending content in the long tail.

Erik Schmidt is a Senior Scientist at Pandora. Most recently, he led the development of Thumbprint Radio, a hyper-personalized product that has attracted over 20 million listeners to date. He has contributed numerous recommendation strategies to the Pandora ecosystem spanning playlists, station discovery, and concerts. Before joining Pandora, Schmidt was a postdoctoral researcher in the Music and Entertainment Technology Laboratory (MET-lab) at Drexel University in Philadelphia. His general research interests lie in the areas of machine learning, recommendation systems, and digital signal processing.

Is Your Organization Ready for ML?

As data science has become increasingly popular, many organizations rush to hire ML experts without laying the proper foundation to ensure their success, including creating proper database architecture, building out essential data science technology, establishing data governance, and instilling data-driven decision-making throughout the organization. Absent these elements, many ML experts join companies excited to deploy their data science expertise only to end up marred in data cleaning or lobbying for tech resources. In this presentation, I discuss how organizations can prepare their organization for success, as well as how candidates can diagnose whether the organization is truly ready for ML.

Dr. Gershkoff is Chief Data Officer for Ancestry, which specializes in genealogy and consumer genomics. Previously, she was Chief Data Officer at Zynga. During her career, she has led the Customer Analytics & Insights team at eBay, served as Chief Data Scientist at WPP, and was Head of Media Planning at Obama for America, where she designed the campaign’s advertising and analytics strategy. Gershkoff was named one of America’s “40 Under 40” leading entrepreneurs, one of the Top 50 Women to Watch in Tech, and one of San Francisco's Most Influential Women in Business. She holds a Ph.D. from Princeton University.