AI Superstream: NLP in Production

Schedule

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

Antje Barth: Introduction (5 minutes) - 8:00am PT | 11:00am ET | 3:00pm UTC/GMT

• Antje Barth welcomes you to the AI Superstream.

Antje Barth: Keynote—The Top 5 Trends in NLP in 2022 (10 minutes) - 8:05am PT | 11:05am ET | 3:05pm UTC/GMT

• In her opening keynote, Antje Barth discusses the top five trends and use cases in NLP in 2022.

Thom Ives: Keynote—A Math Machine That Skipped School and Learns on the Job (20 minutes) 8:15am PT | 11:15am ET | 3:15pm UTC/GMT

• You have a large and growing corpus of documents. Categorization of new documents must be done fast! Each new document greatly helps the categorization process. But waiting for math machine retraining will impede the speed. So don't retrain. Don't even train. Thom Ives explains how to use principles common to word vectorization machines and basic linear algebra techniques to avoid training.
• Thom Ives is the founder of Integrated Machine Learning & AI, a large and growing group of data scientists seeking to learn and grow more together. He’s also senior data scientist at Echo Global Logistics. And his 42,000+-strong following on LinkedIn is growing fast through appreciation of his posts on data science. Over his long career, Thom has developed a wide range of analytical models using data, multiphysics, and experiments. While he loves predictive modeling, he's equally passionate about automating the entire data pipeline to achieve the greatest return on data. Thom is married, has nine kids, and lives in Eagle, Idaho.

Hamid Shojanazeri: How to Deploy Low Latency NLP Models with PyTorch (30 minutes) - 8:35am PT | 11:35am ET | 3:35pm UTC/GMT

• Join Hamid Shojanazeri to explore modern NLP models in production and learn how Meta and other industry players have adopted them to address some of the biggest challenges in this space. Along the way, you’ll discover some of the tools and practices offered by PyTorch and its ecosystem and examine the challenges and available solutions around model optimization and serving NLP models in production, including optimizations for various hardware choices such IPEX, TensorRT, and ONNX Runtime. Hamid will also offer an overview of TorchServe, a model serving solution offered by PyTorch, and its features for robust deployment in production, along with its integrations with other open source and managed services in the production space.
• Hamid Shojanazeri is a partner engineer at PyTorch who works on OSS high-performance model optimization and serving. Hamid holds a PhD in computer vision. Previously, he was a researcher in multimedia labs in Australia and Malaysia and an NLP lead at Opus.ai. He enjoys working on AI OSS tooling and is also an art enthusiast.
• Break (10 minutes)

Surbhi Rathore: How to Build a Startup with Conversation Intelligence (30 minutes) - 9:15am PT | 12:15pm ET | 4:15pm UTC/GMT

• There’s been massive investment in new startups and products built to capitalize on the abundance of conversation data flowing across communication tools and products. Drawing on her experience at Symbl.ai, Surbhi Rathore offers a perspective on building a venture-backed startup and bridging the understanding gap for creating this new category, focusing on value generation and ROI from audio or video conversations. You’ll learn about the customer discovery process, how to set up a hypothesis and initial market segments for product-market fit, and how to measure iterative value with the platform and repeatable use cases. You’ll also explore solutions and early impact in specific industries like recruiting, events platforms, meeting products, sales, marketing, and customer experience.
• Surbhi Rathore is on a mission to democratize contextual AI by making it simple to deploy conversational intelligence for any stage software. She’s cofounder and CEO of Symbl.ai—a company empowering developers and businesses to enable AI in their voice and video conversations at scale, without the need for in-house data science expertise—where she’s bringing to life her vision for a programmable platform built on easy-to-deploy APIs.

David Talby: Automated Medical Question Answering in Practice (30 minutes) - 9:45am PT | 12:45pm ET | 4:45pm UTC/GMT

• Recent advances in natural language processing have enabled material gains in accuracy on the academic benchmarks of reading comprehension and question answering. However, real-world use cases pose challenges beyond what current models address: handling domain-specific jargon, inferring answers from multiple sources, dealing with uncertain or conflicting information, and answering questions at varying levels of specificity. David Talby shares lessons learned from two real-world healthcare case studies—one with a major US healthcare system on answering questions about clinical guidelines and the other with one of the world’s largest pharmaceutical companies on answering questions about patient histories.
• David Talby is a chief technology officer at John Snow Labs, helping healthcare and life science companies put AI to good use. David is the creator of Spark NLP, the world’s most widely used natural language processing library in the enterprise. He has extensive experience building and running web-scale software platforms and teams—in startups, for Microsoft Bing in the US and Europe, and to scale Amazon’s financial systems in Seattle and the UK. David holds a PhD in computer science and master’s degrees in both computer science and business administration.
• Break (5 minutes)

Braden Hancock: Best Practices for Automated Data Labeling in NLP (Sponsored by Snorkel AI) (30 minutes) - 10:20am PT | 1:20pm ET | 5:20pm UTC/GMT

• Labeling training data is labor-intensive and exhausting—making it one of the biggest bottlenecks AI teams face today. To alleviate this pain point, practitioners have sought ways to automate the labeling process. But automate too little (e.g., with manual labeling optimizations such as active learning or model-assisted labeling) and the gains are marginal. Automate too much and your model becomes disconnected from the essential human-provided domain knowledge it needs. Join Braden Hancock to discover the key to truly transformative (10x to 100x) efficiency improvements: changing the interface to labeling altogether, moving from manual labeling (collecting individual labels one by one) to programmatic labeling with labeling functions that capture labeling rationales. The result is a labeling process that’s significantly more scalable, adaptable, and governable.
• Braden Hancock is a cofounder and head of technology at Snorkel AI. Previously, Braden researched and developed new interfaces for machine learning systems in academia (at Stanford, MIT, Johns Hopkins, and BYU) and industry (at Facebook and Google).
• This session will be followed by a 30-minute Q&A in a breakout room. Stop by if you have more questions for Braden.

Lewis Tunstall: Accelerating Transformers with Hugging Face Optimum (30 minutes) - 10:50am PT | 1:50pm ET | 5:50pm UTC/GMT

• Since their introduction in 2017, transformers have become the de facto standard for tackling a wide range of NLP tasks in both academia and industry. However, in many situations accuracy isn’t enough—your fancy model isn’t very useful if it’s too slow or too large to meet your application’s business requirements. Lewis Tunstall explains how to accelerate transformers with Hugging Face Optimum, a new open source library that enables developers to train and run these models on targeted hardware.
• Lewis Tunstall is a machine learning engineer at Hugging Face, where he focuses on developing tools for the NLP community and teaching people to use them effectively. He’s built machine learning applications for startups and enterprises in the domains of NLP, topological data analysis, and time series. Lewis has a PhD in theoretical physics and has held research positions in Australia, the US, and Switzerland.

Antje Barth: Closing Remarks (5 minutes) - 11:20am PT | 2:20pm ET | 6:20pm UTC/GMT

• Antje Barth closes out today’s event.

Upcoming AI Superstream events:

• MLOps - December 7, 2022