Moderator Introduction and Housekeeping

Speakers: Emily Webber, Machine Learning Specialist Solutions Architect, AWS | Neel Kishan, Business Development Manager AI/ML, AWS 

GPT-3, Wu Dao, Megatron-LM, Compositional Transformers and more. In this session we'll explore a variety of extreme scale NLP models and the distributed science and architecture that make them tick. In this session you'll learn about the newly launched distributed training toolkits on the AWS cloud and learn about how customers from Hyundai to Hugging Face are pushing the boundaries on transformer-based models. You'll learn about custom hardware AWS is developing to lower costs and boost performance. You'll also learn about how to partner with AWS to bring your unique technology to the enterprise cloud market!

Moderator: Ari Kalfayan, Senior Business Development Manager - AI/ML, AWS
Speakers: Tamás Görbe, PhD Co-founder and COO, Menten AI | William Falcon, Co-founder and CEO, Grid.ai | Matthew Rocklin, CEO, Coiled Computing | Chris Van Pelt, Co-founder, Weights & Biases

Startups focused on artificial intelligence/machine learning (AI/ML) will face a variety of challenges as they begin building out their product to scale. Conditions must be right, and data centers are not a viable option. Instead, you must have a cloud provider in place to facilitate. To help learn from those who’ve done it before, we’ve gathered AI/ML founders from some of the world’s top startups to give a peek behind the scenes into the secrets of their own massive scaling success. We’ll get the inside scoop on their untold stories and learn how they persevered through unexpected difficulties. You'll learn about common mistakes and get practical advice to help you navigate the challenges ahead.

Speaker: Rob Ferguson, Principal Business Development Manager - AI/ML, AWS

Scaling up from a single engineer working off of their laptop to a dedicated team is an exciting milestone. But with growth comes growing pains. As you scale up your machine learning (ML) team, it's essential to leverage cloud services and tools just like you do for the rest of your development teams. Discover how to set up a data lake and implement it into an ML experiment workflow, how to prepare an end-to-end workflow to easily share the workload, and other tips for scaling your startup.

Speakers: Sundar Ranganathan, Global Head of Business Development - ML Frameworks, AWS | Geeta Chauhan, PyTorch Partner Engineering Head, Facebook

PyTorch, the popular open-source ML framework, has continued to evolve rapidly since the introduction of PyTorch 1.0, which brought an accelerated workflow from research to production. We’ll deep dive into some of the most important new advances, including large scale distributed training, techniques for performance optimization, production scale deployments along with best practices for building Responsible AI solutions.

MLOps Workshops hosted by Tecton.ai, OctoML, Fiddler, Pachyderm, and Activeloop.

Speaker: Achal Shah, Software Engineer, Tecton.ai

Feature stores play a pivotal role in the modern machine learning stack. More and more data scientists and engineers are working together to create and manage features for both model training and for real-time inference. But how do you build, deploy, and use a feature store in the first place?In the tutorial, we will walk through a use case to build a real-time credit scoring application using Feast and AWS storage components: Redshift (offline store) and DynamoDB (online store). In particular, we will talk through how to:

• Create a training dataset as a loan table, which holds historical loan data with accompanying features, including a target variable: whether a user has defaulted on their loan.
• Demonstrate on-demand feature transformations
• Build a predictive model and use SageMaker for model experiments
• Serve real-time predictions with this model by using DynamoDB as the online feature store

Speaker: Phil Mazenett, Senior Solutions Engineer, OctoML

The 2021 Enterprise Trends in Machine Learning survey revealed that the cost of deploying ML to production is skyrocketing and that the time to deploy has increased year over year. This session is for organizations looking to evolve their ML deployment workflows and processes to accelerate time to production. Attendees will learn about new model performance optimization techniques and open source technologies. As well, they will get guidance on how to benchmark performance of optimized models across a variety of hardware options. Attendees will also take away strategies for automating manual steps in their deployment pipeline. And finally this workshop will cover actionable insight on how these approaches can be implemented to work with your existing AI/ML implementations on AWS.

Speakers: Amy Hodler, Lead Evangelist, Fiddler | Amal Iyer, Data Scientist, Fiddler

ML models tend to lose their predictive power over time and can fail silently. In this session, we’ll review how to identify and stay ahead of the common culprits: model drift, data integrity, outliers and bias. You’ll see how cutting-edge explainable AI and model analytics can quickly find the root cause of operational issues. And we’ll outline how model and cohort comparison help teams iterate and get new models in production faster.

We’ll demonstrate how leading enterprises on AWS are achieving trustworthy AI by integrating model performance management into their MLOps lifecycle, with Amazon SageMaker and AWS. You’ll walk away knowing how to use continuous ML monitoring and explainability to achieve optimal model performance and accelerate business outcomes.

Speaker: Jimmy Whitaker, Senior Machine Learning Developer and Developer Advocate, Pachyderm, Inc.

Natural language processing (NLP) is one of the most difficult areas in machine learning. Language is always changing because how we communicate is always changing. When it comes to machine learning our models are only as effective as the data they’ve been trained on. This is why any NLP task ultimately becomes an exercise in data-driven development.

Managing data for machine learning, however, has always been an afterthought. Tracking, versioning, and testing are usually introduced after it’s too late. In this workshop, we’ll show you how to manage your data, automate your training pipelines, and shift to a data-centered development process on a financial market sentiment model, and run it all deployed on AWS with EKS and S3.

In this workshop, attendees will learn how to:

• Train a market sentiment model
• Automate your NLP pipelines
• Shift to data driven development with Pachyderm and AWS

Speakers: Davit Buniatyan, CEO and Founder, Activeloop | Ivo Stranic, Head of Product, Activeloop

Learn how building a solid data foundation can help unlock the full potential of SageMaker and make training more accurate and cost-effective for computer vision models. In this session, we present Activeloop Hub - a simple, open-source API for transforming and streaming data while training models at scale. We discuss how adopting a dataset format specifically designed for AI allows for easy creation, storing, version-control and collaboration for CV datasets of any size. In addition, we present an instant way to visualize and explore your data. Overall, you will be able to adopt a radically simpler SageMaker experience - both for you, and your computer vision datasets.

Moderator closing remarks