Skip to content
O'Reilly home
Artificial Intelligence

AI Superstream Series: Scaling AI

Published by O'Reilly Media, Inc.



September 22, 2021

4:00 p.m. - 8:00 p.m. Coordinated Universal Time

This event has ended.

What you’ll learn and how you can apply it

  • Understand what scaling AI does (and doesn’t) include
  • See what scaling AI might look like from design through deployment
  • Explore what current AI leaders are achieving through scaling
  • Discover real-world technical applications

This live event is for you because…

  • You're a machine learning engineer or data scientist interested in the challenges and benefits of scaling.
  • You’re interested in learning how industry experts handle scaling.
  • You're wondering how to improve your own AI and machine learning.


  • Come with your questions
  • Have a pen and paper handy to capture notes, insights, and inspiration

Recommended follow-up:


The timeframes are only estimates and may vary according to how the class is progressing.

Antje Barth: Introduction (5 minutes) - 9:00am PT | 12:00pm ET | 4:00pm UTC/GMT

  • Antje Barth welcomes you to the AI Superstream.

Antje Barth: Keynote—Think Big. Start Small. Scale Fast. (15 minutes) - 9:05am PT | 12:05pm ET | 4:05pm UTC/GMT

  • “Think big. Start small. Scale fast.” This famous piece of business advice doesn’t only apply to startups; it applies equally to your AI projects. As you move from idea to production, you need to process growing amounts of data, train models of all sizes, automate manual steps in the machine learning development workflow, and scale your AI application to meet your inference demand. In her keynote address, Antje Barth discusses the challenges of scaling AI projects and shares strategies and tips that will help you scale fast and efficiently.

Geoff Horrell: Keynote—NLP at Scale in Financial Services (Sponsored by LSEG Labs) (10 minutes) - 9:20am PT | 12:20pm ET | 4:20pm UTC/GMT

  • Major advances in NLP have been achieved over the last few years thanks to increases in processing power, data availability, open source, and new techniques. NLP is already being used across financial services for many use cases, like sentiment analysis and recommendation systems, but it has yet to scale. A new industry survey by LSEG Labs, being shared for the first time at Scaling AI, reveals which NLP tools are maturing, how technology and data science skills have evolved, and what’s holding financial services firms back from executing and scaling in the new age of machine learning. Following this keynote, members of LSEG Labs will be on hand to share a range of live projects, along with their practical experience applying NLP at scale using news, transcripts, and research data.
  • Geoff Horrell is the group head of innovation at LSEG. Over the last two decades, he has created and launched numerous financial technology and data products. Geoff manages LSEG Labs, a corporate innovation team of engineers, data scientists, and user experience experts tackling the next generation of challenges in financial markets.
  • This session will be followed by a 30-minute breakout room. Stop by if you have more questions for LSEG Labs.

Venkatesh Ramanathan: Scaling Optimization of Graph Neural Networks (Sponsored by Intel) (30 minutes) - 9:30am PT | 12:30pm ET | 4:30pm UTC/GMT

  • To efficiently service its millions of global customers, PayPal has developed novel approaches to graph neural networks that operate at a massive scale. Venkatesh Ramanathan shares a variety of challenges he encountered and addressed when developing and optimizing these graph neural networks, with compute, orchestration, and experimentation as key focus areas. You’ll learn how he ran hyperparameter search in parallel across significant compute width and how tools offered by SigOpt, an Intel acquisition, made this task much easier and more sample efficient. If you’re considering graph neural networks, this is the talk for you.
  • Venkatesh Ramanathan is a director of data science at PayPal, where he leads several applied research initiatives, including ML on graphs. An expert in AI/ML and big data technologies, he has over 25+ years’ experience designing, developing, and leading teams to build scalable server-side software and doing AI/ML research, and he’s worked on various problems in the areas of anti-spam, phishing detection, and face recognition. Venkatesh holds a PhD in computer science with specialization in machine learning and natural language processing (NLP).
  • This session will be followed by a 30-minute Q&A in a breakout room. Stop by if you have more questions for Venkatesh.

Adi Polak: Demystifying Scalable Machine Learning with the Spark Ecosystem (30 minutes) - 10:00am PT | 1:00pm ET | 5:00pm UTC/GMT

  • To create good products that leverage AI, you need to run machine learning algorithms on massive amounts of data. Distributed machine learning frameworks, such as Spark ML, help simplify the development and use of large-scale machine learning. With Apache Spark libraries, you can cover the entire basic machine learning workflow, from loading and preparing data to extracting features to fitting the model to scoring. Software and data engineers must understand the workflow in order to leverage what already exists and create more-enhanced products; likewise, tech leads and architects need to understand the workflow and available options to build better architecture and software. Join Adi Polak to explore the end-to-end machine learning workflow and learn how to build one with Apache Spark.
  • Adi Polak is a senior software engineer and developer advocate in the Azure engineering organization at Microsoft. Her work focuses on distributed systems, big data analysis, and machine learning pipelines. In her advocacy work, she brings her vast industry research and engineering experience to bear in educating teams and helping them design, architect, and build cost-effective software and infrastructure solutions that emphasize scalability, team expertise, and business goals. Adi is a frequent presenter at worldwide conferences, a training course instructor, and the author of Machine Learning with Apache Spark (forthcoming from O’Reilly). When she isn't building machine learning pipelines or thinking up new software architecture, you can find her on the local cultural scene or at the beach.
  • Break (10 minutes)

Victor Dibia: NeuralQA—A Usable Library for Question Answering on Large Datasets Using BERT-Based Models (30 minutes) - 10:40am PT | 1:40pm ET | 5:40pm UTC/GMT

  • The ability to provide exact answers to queries framed as natural language questions can significantly improve the user experience in many real-world applications. Join Victor Dibia for a deep dive into question answering that will help you get going. You’ll examine methods such as the traditional two-stage pipeline and more recent approaches like dense neural retriever and reader, then get an overview of NeuralQA, an easy-to-use open source library for neural question answering. Victor will show you how NeuralQA works as he takes you through its key features, including contextual query expansion, the configurable visual interface for search, and support for model explanations. Along the way, you’ll explore the failure modes of QA systems and learn practical strategies for improving them, grounded in a use case for question answering on legal documents.
  • Victor Dibia is a machine learning research engineer at Cloudera Fast Forward Labs. His interests lie at the intersection of applied machine learning, human-computer interaction (HCI), and computational social science. His research has been published at conferences such as EMNLP, AAAI, and CHI, receiving multiple best paper awards. Victor’s an IEEE senior member, a Google Certified Professional Data Engineer and Cloud Architect, and a Google Developers Expert in machine learning. Previously, he was a research staff member at IBM Research, a technical lead at MIT Global Startup Labs, a researcher at the Innovation Management Lab at Athens Information Technology, and the founder and lead developer for a small startup focused on West African markets.

Spence Green: Scaling Localization with AI and Automation (Sponsored by Intel) (30 minutes) - 11:10am PT | 2:10pm ET | 6:10pm UTC/GMT

  • As the world becomes more digital and more globally connected, companies increasingly face the daunting task of creating and managing global customer experiences for multilingual customer bases. Spence Green explains how AI and automation are redefining the speed, simplicity, and scale with which businesses can accomplish these tasks, making it easier to identify gaps in the customer experience and close them with high-quality localized content and interactions. You’ll learn how to scale your global experience program, drive higher ROI with AI, reduce complexity with automation, and more.
  • Spence Green is the cofounder and CEO at Lilt. Previously, he was a fellow at XSeed Capital; worked in software and research at Google, where he worked on Google Translate and developed a shallow syntactic language model for improving English-to-Arabic translation; and was a technical lead, project manager, and software engineer at Northrop Grumman, where he worked on large projects including a national air defense system and avionics packages for naval aircraft. Spence graduated with highest distinction from the University of Virginia with a bachelor’s degree in computer engineering and holds both a master’s degree with a distinction in research and a PhD in computer science from Stanford University.
  • This session will be followed by a 30-minute Q&A in a breakout room. Stop by if you have more questions for Spence.
  • Break (10 minutes)

Rebecca Bilbro: The Promise and Peril of Very Big Models (30 minutes) - 11:50am PT | 2:50pm ET | 6:50pm UTC/GMT

  • In the machine learning community, we're trained to think of size as inversely proportional to bias, driving us to ever larger datasets, increasingly complex model architectures, and ever better accuracy scores. But bigger doesn't always mean better. Rebecca Bilbro investigates key challenges that arise when scaling AI and shares practical solutions. What data quality issues emerge in large datasets? What complications surface as features become more geodistributed (e.g., diurnal patterns, seasonal variations, datetime formatting, multilingual text, etc.)? What happens as models attempt to extrapolate bigger and bigger patterns? Why is it that the pursuit of megamodels has driven a wedge between the ML definition of “bias” and the more colloquial sense of the word? Perhaps the time has come to move away from monolithic models that reduce rich variations and complexities to a simple argmax on the output layer and instead embrace a new generation of model architectures that are just as organic and diverse as the data they seek to encode.
  • Dr. Rebecca Bilbro is a data scientist, Python and Go programmer, and applied NLP developer who has worked at numerous startups in fields ranging from the public sector to media and entertainment to enterprise security. As founder and CTO of intelligent distributed systems company Rotational Labs, Rebecca specializes in machine learning optimization and API development in distributed data systems. She’s also cocreator of the popular open source ML library Yellowbrick, a faculty member in Georgetown University's Data Science Certificate Program, and author of Applied Text Analysis with Python. Rebecca holds a PhD from the University of Illinois Urbana-Champaign, where her research centered on communication and visualization in engineering, and a bachelor’s degree in mathematics and English from Skidmore College.

Robert Nishihara: Scalable Python and Machine Learning with Ray (30 minutes) - 12:20pm PT | 3:20pm ET | 7:20pm UTC/GMT

  • Distributed computing is becoming the norm, driven in large part by machine learning applications. However, building distributed applications today requires tons of expertise—putting it out of reach for the vast majority of developers. Robert Nishihara explains how Ray aims to make programming a cluster of machines as easy as programming a laptop, enabling many more developers to take advantage of the advances in machine learning to solve harder problems without needing to be experts in distributed systems.
  • Robert Nishihara is one of the creators of Ray, a distributed system for scaling Python and machine learning applications. He’s one of the cofounders and CEO of Anyscale, the company behind Ray. He did his PhD in machine learning and distributed systems in the Computer Science Department at UC Berkeley and majored in math at Harvard.

Antje Barth: Closing Remarks (5 minutes) - 12:50pm PT | 3:50pm ET | 7:50pm UTC/GMT

  • Antje Barth closes out today’s event.

Upcoming AI Superstream events:

  • Securing AI - December 1, 2021

Your Host

  • Antje Barth

    Antje Barth is a senior developer advocate for AI and machine learning at AWS. She’s the coauthor of Data Science on AWS and frequently speaks at AI and machine learning conferences, online events, and meetups around the world. Antje is also passionate about helping developers leverage big data, container, and Kubernetes platforms in the context of AI and machine learning. She’s cofounder of the Düsseldorf chapter of Women in Big Data.

Start your free 10-day trial

Get started

Want to learn more at events like these?

Get full access to O'Reilly online learning for 10 days—free.

  • checkmark50k+ videos, live online training, learning paths, books, and more.
  • checkmarkBuild playlists of content to share with friends and colleagues.
  • checkmarkLearn anywhere with our iOS and Android apps.
Start Free TrialNo credit card required.