Applied AI Summit Applied AI Summit schedule

Sarah is a Partner at Amplify Partners where she focuses on early-stage investments in machine intelligence, data science, and data management. Sarah has several years of experience in developing data acquisition strategies and leading machine and deep learning-enabled product development. As head of data at Mattermark, she led a team to collect and organize information on over one million private companies; as a consultant at Palantir and as an analyst at Cyveillance, she implemented analytics solutions for municipal and federal agencies; and as a program manager at the Center for Advanced Defense Studies, she directed projects on adversary behavioral modeling and Somali pirate network analysis. Sarah earned a B.A. in International Security Studies from Stanford University.

Eddan Katz is a global expert in technology law and policy, a digital rights activist, and access to knowledge advocate. Now working with the Center for AI and Digital Policy on the AI and Democratic Values report, he developed the pilot project methodology at the World Economic Forum’s Centre for the Fourth Industrial Revolution network, and led multi-stakeholder initiatives on the AI/ML and Data Policy platforms. He was the first Executive Director of the Yale Information Society Project, the International Affairs Director at the Electronic Frontier Foundation and co-founder of the Sudo Room hackerspace in downtown Oakland.

Frankie Cancino is a Data Scientist at Mercedes-Benz Research & Development, working on applied machine learning initiatives. Prior to joining Mercedes-Benz R&D, Frankie was a Senior AI Scientist at Target AI, focused on methods to improve demand forecasting and anomaly detection. He is also the organizer and founder of Data Science Minneapolis. Data Science Minneapolis is a community that brings together professionals, researchers, data scientists, and AI enthusiasts.

Enhance Recommendations in Uber Eats with Graph Convolutional Networks

Uber Eats has become synonymous to online food ordering. With increasing selection of restaurants and dishes in the app, personalization is quite crucial to drive growth. One aspect of personalization is better recommendation of restaurants and dishes to the users so they can get the right food at the right time.

In this talk, we present how to augment the ranking models with better representations of users, dishes and restaurants. Specifically, we show how to leverage the graph structure of Uber Eats data to learn node embeddings of various entities using state of the art Graph Convolutional Networks implemented in Tensorflow. We also show that these methods perform better than standard Matrix Factorization approaches for our use case.

Key Takeaway - The audience will learn about how to build deep learning models on graph data using Graph Convolutional Networks to obtain better entity representations to use for recommendation. They will also learn about strategies to scale the model to very big datasets.

Biography: Piero Molino is a Senior Research Scientist at Uber AI with focus on machine learning for language and dialogue. Piero completed a PhD on Question Answering at the University of Bari, Italy. Founded QuestionCube, a startup that built a framework for semantic search and QA. Worked for Yahoo Labs in Barcelona on learning to rank, IBM Watson in New York on natural language processing with deep learning and then joined Geometric Intelligence, where he worked on grounded language understanding. After Uber acquired Geometric Intelligence, he became one of the founding members of Uber AI Labs. He currently leads the development of Ludwig, a code-free deep learning framework.

Ankit Jain currently works as a machine learning tech lead at Meta where he works on a variety of growth ranking and business integrity problems. Previously, he was a ML researcher at Uber AI where he worked on application of deep learning methods to different problems ranging from food delivery, fraud detection to self-driving cars. He has co-authored a book on machine learning titled TensorFlow Machine Learning Projects. Additionally, he’s been a featured speaker and published papers in many of the top AI conferences and universities. He was recently awarded as top 40under40 data scientists from India. He earned his MS from UC Berkeley and BS from IIT Bombay (India).

In-Storage Distributed Machine Learning for the Edge

Cloud-only architectures will soon not be able to keep up with the volume and velocity of data across the network, therefore gradually reducing the value that can be created from these investments. Edge computing can help solve the limitations in current infrastructure to enable mission-critical, data-dense IoT and other advanced digital use cases by reducing or eliminating data movement and address latency and energy efficiency bottlenecks. To address the problems above in the context of ML applications, it is necessary to perform training and inference at the edge, transmitting only processed data (metadata) or full data only when necessary. Doing this, however, faces the limitation that most devices do not present strong computing capabilities and, even if they did, it would take too much energy to make them work.

Big data analytics solutions, such as Hadoop, have addressed the performance challenge by using a distributed architecture based on a new paradigm that relies on moving computation closer to data. Similarly, by pushing the “move computation to data” paradigm to its ultimate limit we enable highly efficient and flexible in-storage processing capability in solid state drives, i.e., computational storage. By moving data processing tasks closer to where the data resides, we dramatically reduce the storage bandwidth bottleneck, data movement cost, and improve the overall energy efficiency creating an ideal platform for Machine Learning at the Edge.

NGD’s computational storage device (CSD) provides a seamless programming model based on a Linux OS and high-level programming languages thanks to a complete standard network software and protocol stack. It is the first commercially available SSD that can be configured to run a server-like operating system (e.g., Linux), allowing general application developers to fully leverage existing tools and libraries to minimize the effort to create and maintain applications running in-storage.

This paper proposes a framework for distributed, in-storage training of neural networks on heterogeneous clusters of computational storage devices. Such devices contain multi-core application processors as well SIMD engines and virtually eliminate data movement between the host and storage, resulting in both improved performance and power savings. More importantly, this in-storage processing style of training ensures that private data never leaves the storage while fully controlling the sharing of public data. Experimental results have shown up to 2.7x speedup and 69% reduction in energy consumption and no significant loss in accuracy.

Vladimir Alves obtained his PhD degree in Microelectronics when the 500nm CMOS process was all the hype. Since then Vladimir worked in the academia, startups and multinational companies architecting and developing System on Chips. In the last 15 years he has been focusing on solid state storage technology and is now the co-founder and Chief Technology Officer at NGD Systems helping create innovative technology that pushes the boundaries of storage and computation.

Sentilink uses machine learning to stop synthetic fraud. Specifically, the Sentilink API allows lenders to determine in real time whether someone applying for their loan is a real person, or whether they are merely a "synthetic identity" created by a fraudster. In this talk, Sentilink data scientist Seth Weidman will discuss the subtle signals Sentilink uses to detect such fraud, and also share lessons we've learned about managing real time machine learning models in production.

Seth Weidman is a data scientist at Sentilink, where he works on their algorithms to stop synthetic fraud. He has done many jobs in the data analytics and machine learning space, from working as a business analyst in management consulting to doing machine learning engineering at Facebook, and recently published an introductory book on Deep Learning with O'Reilly. He has degrees in mathematics and economics from the University of Chicago.

Duolingo’s Growth Model: A framework for understanding, exploration, and forecasting

Duolingo's evolution has required us to have a more fine-grained understanding of the levers that drive user growth. Our Growth Model is a framework that allows us to better characterize our topline DAU and MAU metrics and model how they change over time. Using this framework, we have identified new strategies to unlock user growth, built a robust forecasting engine, and explored new opportunities for product development.

Erin Gustafson is a Senior Data Scientist at Duolingo, where she works on product-focused exploratory modeling, forecasting, and experimentation. She has experience working in Growth and Monetization, where she applies her statistics and ML skills to improving user/product understanding and making data-informed recommendations on future product development. Before joining Duolingo, Erin completed her PhD in linguistics and conducted research on bilingualism.

Personalizing the Online Grocery Substitution Experience

In this talk, I will present a broad overview of personalizing recommendations in Online Grocery. Unlike typical recommender systems used in e-commerce that have roots in collaborative filtering, recommending groceries to customers poses unique challenges. How do we handle periodicity in purchases? How do we blend periodicity in purchases with a broad prior based on seasonality? Should recommendations be tailored at an item level, or more broadly at basket level? I will provide a high level end to end overview of online grocery shopping, and as a case study, I will provide deep insights into personalized substitutions. I will describe our state of the art neural network based model that allows order pickers to substitute out of stock items during order fulfilment. Perhaps in a system where we strive to optimize substitutions for our customers, how do we model the inherent bias, akin to a noisy channel, introduced by order pickers? Apart from providing details on personalized substitutions, I will also expand on recent work that touches on basket level substitutions as a multi- objective problem that takes in to account objectives such as cost control and recipe completion.

Kamiya Motwani is a Staff Data Scientist and manager at Walmart Labs India. She is currently a data science lead in Personalization Team. She has also worked extensively on click prediction for advertisements and has rich practical experience building machines that learn from data. Prior to Walmart Labs, she has worked in prestigious organizations such as Oracle corporation and Yahoo Inc. She holds a Master's degree in Computer science from the University of Wisconsin Madison where she focused extensively on Machine learning and probabilistic modelling. Kamiya has also filed several patents in the area of recommender systems, and published papers at premier conferences including NIPS and IEEE ICASSP.

Zero-To-Hero: Solving the NLP Cold Start Problem

Mailchimp is the world's largest marketing automation platform. Over a billion emails are sent everyday by users through the platform. This mass of marketing text data creates lots of opportunities to leverage natural language processing to improve and create content for users. Like many natural language processing (NLP) practitioners, data scientists at Mailchimp have found annotating text data to be costly, time consuming, and in some cases legally prohibited. So how do they work around it? We'll do a deep dive into how Mailchimp uses state-of-the-art NLP models and unlabeled data to cold start NLP products. We'll cover its data-centric (over model-centric) approach and how it positions its products to facilitate a data flywheel.

Muhammed Ahmed is a Senior Data Scientist at Mailchimp who specializes in natural language processing and computer vision. At Mailchimp, he has majorly contributed to the implementation and deployment of several AI-assisted products including multimodal classification, preview text generation, stock photo recommendation, campaign engagement scoring, and semi-supervised topic clustering using large transformer models (T5, BART, RoBERTa, UNITER, and similar). Most recently, his focus has been on developing a systematic approach to use zero-shot learning to extract arbitrary information from text.

Understanding Text on Images at Scale

Understanding text that appears on images in social media platforms is important not just for improving experiences such as the incorporation of text into screen readers for the visually impaired, but they also help keep the community safe by proactively identify inappropriate or harmful content in a way that pure object detection or NLP systems alone cannot.

This talk describes the challenges behind building an industry-scale scene-text extraction system at Facebook that processes over 2 billion images each day. I'll cover the Deep Learning methods behind building models that perform detection of text in arbitrary orientations with high-accuracy, and how simple convolutional models work extremely well for recognizing text in over 50 languages. A critical aspect of the work is scaling up these models for efficient server-side inference. I'll dive into quantization methods to run neural networks with 8-bit integer weights and activations instead of 32-bit floating points, and the challenges involved in bridging the accuracy gap.

I’m a Researcher at Facebook AI Research working on machine learning for systems where I’m currently exploring reinforcement learning to improve the performance of computer networks. Prior to that, I was part of Facebook AI Applied Computer Vision Research group where I founded and lead the Rosetta project—a large-scale machine learning system for understanding text in images and videos. I had also made extensive improvements to the low-level performance and efficiency of Computer Vision models in production.

Deep Neural Networks for Search and Recommendation Systems at LinkedIn

Deep neural networks like convolutional neural networks (CNN), recurrent neural networks (RNN), and attention-based encoder-decoder networks have made a big impact in several natural language processing (NLP) applications, such as sentence classification, part of speech tagging, and machine translation. In recent years, models like BERT and its variants have improved the state of the art in NLP through contextual word embeddings, and sentence embeddings. Another attraction of these models is that they can be finetuned for target applications.

In this talk, I will describe how we have successfully used deep neural networks for natural language processing and understanding at LinkedIn. In particular, I will discuss our work in query and document understanding, as well as document ranking for search and recommendation systems.

Ananth Sankar is a Principal Staff Engineer in the Artificial Intelligence group at LinkedIn, where he works on multimedia content understanding and natural language processing. During his career, he has also made many R&D contributions in the area of speech recognition. He has taught courses at Stanford and UCLA, given several invited talks, co-authored more than 50 refereed publications, and has 10 accepted patents.

Using Machine Learning for Risk Management within an Insurance Company

Solvency capital requirement (SCR) is the amount of funds that insurance and reinsurance companies operating within the EU are required to hold. The SCR is set at a level that ensures that insurers and reinsurers can meet their obligations to policyholders and beneficiaries over the following 12 months with a 99.5 percent probability. To calculate required capital, large number of possible future scenarios must be simulated which requires large amount of computing power and time. In this talk we highlight how Machine Learning can be used to make these calculation estimates at a fraction of time and resources providing significant value to multiple stakeholders within risk management, investments and product development teams.

Rakesh Rana works as Lead Data Scientist at Nordea Life & Pensions, Sweden. His work focuses on applying AI to solve business problems and create customer value. Rakesh received his M.S. degree in Finance and PhD in Computer Science from Chalmers/University of Gothenburg, Sweden. His work and research interests revolve around using data science and machine learning algorithms mainly within the financial domain.

Generating the Best Game Experience through AI

Happy Elements is the producer of one of the largest active mobile games worldwide. Within our AI lab, we aim to optimize the gameplay experience of each player individually at truly massive scale. Towards this goal, we research and develop machine learning algorithms and systems for dynamic game adaptation. This talk elaborates on how we achieve real-time game content personalization by leveraging high volumes of data through deep contextual bandits, as well as our current research projects in applied deep reinforcement learning.

Key Takeaways: Optimizing game design through AI can improve user LTV/retention and enhance player experience.

Rein Houthooft leads Happy Elements AI Team. Originally from Belgium (EU), Rein received his PhD in EECS from Ghent University. Part of his research was conducted as a researcher at OpenAI and at the Berkeley AI Research lab of UC Berkeley, with a focus deep reinforcement learning and generative models. Previously, Rein was involved in the organization of the annual NeurIPS Deep RL Workshop.

How Machine Learning Powers On-Demand Logistics At DoorDash

DoorDash has a complex, three-sided, and real-time marketplace that presents many challenging problems where machine learning has a lot of impact. This talk will give an overview of where machine learning is used in DoorDash, and then dive deeper into how we use machine learning to power our logistics engine, which is the system that powers the fulfillment of deliveries. Topics covered will include the vehicle routing problem with trillions of combinations, delivery time predictions for all your favorite restaurants, and our exploration with reinforcement learning for logistics.

Recommendations and Search at Pinterest + Embeddings

Over 300 million users come to Pinterest monthly to discover ideas for their creative outlets through our recommendation and search products. Embeddings are a fundamental technology in these systems, powering the match engine to surface relevant and engaging content. We represent all aspects of the Pinterest ecosystem (pins, users, text, and images) under this common representation, enabling us to learn relationships across entity types to jointly optimize for product goals. Join us as we discuss how recent advancements in computer vision, natural language processing, and graph convolutional neural networks power embeddings and together enabled both new product experiences such as Pinterest Lens and improved performance of our core recommendation systems. We also discuss key infrastructure challenges and our solutions to scaling embedding search to web-scale systems.

Andrew is a Staff Software Engineer working in the Visual Search and Applied Science groups at Pinterest. During his career, he was the founding engineer and TL of visual search at Pinterest, leading multiple generations of serving, indexing, and modeling to build products including Pinterest Lens. More recently, he leads the embedding efforts at Pinterest to push the limits of recommendation systems. Andrew received his B.S at UC Berkeley and M.S at Stanford.

Shahmeer Mirza is the Director of R&D at 7-Eleven, where he leads the development of technologies that enable the next-generation of convenience. During his time at 7-Eleven, he has launched multiple projects, including 7-Eleven’s Checkout-Free technology. He is passionate about applied AI and inventing solutions for real-world problems, and has over 20 patents. He was previously at PepsiCo, where he developed advanced automation, computer vision, and machine learning solutions for Industry 4.0 applications. He holds an M.S. in Computer Science and a B.S. in Chemical and Biomolecular Engineering, both from Georgia Tech.

Role of AI in Future Mars Exploration

Shreyansh Daftry is a Research Scientist at NASA Jet Propulsion Laboratory (JPL) in Pasadena, California, working at the intersection of Artificial Intelligence and Space Technology to help develop the next generation of robots for Earth, Mars and beyond. Shreyansh received his M.S. degree in Robotics from the Robotics Institute, Carnegie Mellon University, USA in 2016, and his B.S. degree in Electronics and Communications Engineering in 2013. His research interests spans computer vision, machine learning and autonomous robotics, with a focus on real-time computation, safety and adaptability.