Enterprise AI Summit schedule

AI is a complex science, which has seen incredible advances in theory and implementation of available techniques over the last decade. Most practicing Data Scientists and ML engineers love to use them to solve problems for their own sake. But it takes a lot of discipline to convert a prototype into a business with billions of dollars of revenue.
This talk will take a real world example for taking such a problem from 0 to 1.

- Ready-set-launch: Impact analysis, and stakeholder management
- How to no-go? How to make a tough but fair decision to not launch
- Empowering humans: Building monitoring tools for operations team

Vishal is a Product Leader who has built, launched, and scaled products for marketplaces, logistics, transportation, retail, advertising, search, and gaming, at companies like Amazon, Microsoft, Yahoo, Zynga, Lyft, and Shipt. He has been responsible for $Billions in P&L and opex and is passionate about solving audacious problems with a disciplined, collaborative, and outcome-oriented leadership

Foundation models (also known as Large Language Models or LLMs) are accelerating AI in exciting and very visible ways. These large models are pushing the boundaries of AI primarily thanks to innovation in the ways the datasets they are trained on have been created, rather than through new developments in model architectures. This coincides with a general trend seen in recent years where enterprises are increasingly turning toward data-centric rather than model-centric approaches to ML. In this talk, we explore the connections between these two trends, and how the advent of foundation models has made data-centric AI the only viable path for creating high-quality, specialized models for high-value applications.

Stacking deep learning components is an increasingly attractive way to increase model accuracy - but it often comes at the cost of explainability. AI teams increasingly need to build explainability into their core model architecture or as an explicit post-processing layer. In this session, we’ll discuss why AI teams should care about explainability and then dive into explainability strategies, detailing different solutions tailored for end customers vs operators vs developers. We’ll draw on real examples from Opendoor systems used to buy and sell hundreds of thousands of homes algorithmically.

- How to measure the business value of AI explainability
- How to decide when to invest in AI explainability and when not to
- How to tailor AI explainability strategies to best serve different user types

Sam is passionate about building products at the intersection of finance and machine learning. He is currently the Director of Product for the Pricing and Machine Learning Products at Opendoor, a late-stage startup that uses algorithms to buy and sell homes instantly, saving homeowners the hassle and uncertainty of listing their homes and hosting open houses.

Prior to Opendoor, he was a co-founder and Product Manager at Ansaro, a SaaS startup using data science and machine learning to help companies improve hiring decisions. Earlier in his career, Sam worked at genomics software startup SolveBio, investment firm TPG, and consulting firm Bain & Company. Sam holds degrees in Math and International Relations from Stanford and an MBA from Harvard.

The central piece of Pinterest’s technical stack is our recommendation system, which brings responsive and personalized content from a corpus of over 300 billion pins to more than 400 million users. Join us for a deep dive into this system and learn how we are able to serve millions of recommendations per second at millisecond latencies. We'll share our journey of evolving ML serving systems to incorporate state-of-the-art model components such as transformers, and how we integrated GPU serving into our machine learning serving system to serve 100x bigger models.

- A deep dive into Pinterest's technical stack via their recommender systems

- How to incorporate state-of-the-art model components such as transformers

- How to integrate GPU serving into our machine learning serving system to serve 100x bigger models

Zhiyuan is an engineering manager for ML Serving Platforms at Pinterest. His team builds large-scale machine learning inference systems that serve billions of events per minute and drive applications in ads, search, recommendation, and safety. Previously, Zhiyuan was an early member of visual search and core infrastructure teams at Pinterest, where he worked on various recommendation and async execution systems.

Acubed, Airbus' innovation center in Silicon Valley, is building scalable, certifiable autonomy systems to bring about a significant increase in safety and efficiency in the next generation of commercial aircraft. We are making rapid progress in applying machine learning to develop autonomous flight systems by:

 -Architecting dedicated Machine Learning models

-Creating realistic simulations

-Developing safety critical software and hardware

-Advancing our flight testing systems with a flight test aircraft

-Scaling our data collection capabilities and advancing our algorithms

– Without focusing on cost, what are the key returns from AI applications and what should they be aiming to achieve?

- With use-cases of both successes and failures, what are the common pitfalls when it comes to applying new AI frameworks and methodologies?

- Where should the attention be focused and are there certain lifecycle stages that can help generate further success if implemented correctly?

Girija Narlikar is a Director of Engineering in Ads at Instacart, the leading online grocery platform in North America.  Girija joined Instacart in March 2021, after a 3.5 year stint in Google Ads, where she led teams using Machine Learning to identify sensitive content in text, image and videos, including political or COVID-19 related misinformation.  Previously, she worked at Facebook and co-founded an AI-driven start-up in India.  Girija holds computer science degrees from IIT Bombay and CMU.  Outside of work she enjoys hiking, sports, as well as improvising "healthy" new recipes and ordering esoteric ingredients for them on Instacart!

- Current progress towards building intelligent writing assistants for iterative text revision

- How have intelligent writing assistants shifted in their abilities over recent years?

- How can this progress be applied and where can both industry or society benefit from this?

The machine learning landscape is rapidly evolving, with advancements ranging from task-specific supervised training to fine-tuning large foundational models for various tasks, and from virtual machines on the cloud to managed machine learning solutions on the cloud. Additionally, the computational resources available have progressed from single-core machines to powerful high-bandwidth interconnected A100 nodes. Also, both open source as well as vendor tooling and offerings in this eco system are exploding. To navigate this ever-changing environment, it is important to take a systems thinking approach and have a decision framework in place to make sense of the complexity.

- A systems thinking approach gives a comprehensive view of the machine learning ecosystem instead of focusing on individual tools

- It emphasizes the significance of adaptability and flexibility in the ML landscape.

- By embracing a long-term perspective, it allows to better prepare and take advantage of advancements in different areas.

Sravya works as a Senior Machine Learning Engineer at Adobe, where she is part of the team that develops the Sensei Platform, which enables machine learning capabilities throughout Adobe's product line. The platform is designed to provide consistent and reusable building blocks for Machine learning with a strong emphasis on software engineering. Prior to joining Adobe, Sravya co-founded ImpactAI.org, a non-profit organization that utilizes machine learning to address issues in sectors such as health and disaster recovery. She has experience working in both academic and industrial settings and has expertise in designing data and distributed systems for both machine learning and high-performance computing. Sravya is a strong supporter of open-source software and communities, and previously held the role of Vice President for the Apache Sentry project.

The talk will focus on how Ecommerce search engines like Instacart effectively drive a lot of business impact using machine learning. We will start with an overview of the architecture of a modern Ecommerce search engine that can handle large amounts of product data. Then we will talk about some of the typical challenges faced in showing highly relevant search results to customers, and some of the latest machine learning techniques applied to overcome the challenges. We will also cover some of the operational challenges faced by ML teams when building and productionizing highly performant ML models. The talk will conclude with some of the practical lessons learnt in our journey of scaling up ML models to meet the high level of expectations from our customers.

- How do E-commerce search engines effectively drive business impact?

- What are the typical challenges faced in highly relevant customer searches and how can ML techniques be applied to tackle these?

- How do ML Teams build and productionize highly performant ML models and what are some of the practical lessons learned?

Tejaswi Tenneti is a Director of Machine Learning at Instacart. His team is responsible for building ML models for user intent understanding, neural text modeling and multi-objective ranking for search and recommendation use cases across multiple content types like products and recipes. He graduated with a master's degree in CS from Stanford University with a specialization in Artificial Intelligence. He has over a decade of experience in leading search and discovery efforts across multiple domains like e-commerce, points of interest in maps, job search and enterprise.

Modern personalized video recommender systems optimize for multiple objectives and stakeholders in a real-world system. To do this effectively, such systems utilize multi-task learning models to rank both long and short videos more accurately and provide the most relevant content tailored to users' interests. This talk will focus on recent machine learning trends in personalized video recommendations and how they have emerged as key elements to address various challenges in this space. We will provide a comprehensive overview of different approaches to develop more sophisticated and effective methods to increase user engagement and user satisfaction in large-scale video recommender systems.

- What are the various challenges in building modern personalized video recommender systems?
- How have recent machine learning (ML) trends enabled us to address these challenges?
- What are some recent ML approaches to increase user engagement and user satisfaction in large-scale video recommender systems?

Amey Dharwadker is a Staff Machine Learning Engineer at Facebook. He leads the Video Recommendations Core Ranking team building personalization models for billions of users, primarily using deep learning. He also delivered significant improvements to Facebook's News Feed and Ads ranking models. Prior to that, he developed computer vision based Advanced Driver Assistance System algorithms at Analog Devices. He has an active presence in the academic community and is a member of the program committee for top-tier conferences including AISTATS, ECIR, WWW and peer-reviewed journals specializing in recommender systems, information retrieval and deep learning. He has also served on the juries of international tech competitions, including the Edison Awards.

The open secret of Enterprise Machine Learning across the industry is that ML models are often not built and deployed according to best practices or with state of the art platform tools. For these workflows, over time, innovation becomes frustrating and resources are monopolized by the growing burden of maintenance. Companies with a healthy tech culture know that periodically we must roll-up our sleeves and refactor these parts of our portfolio. This session will share lessons learned in identifying and rehabilitating machine learning pipelines in need of attention.  

- The pressures of getting models to market and the rising bar of engineering standards means we always have machine learning pipelines with tech debt in our portfolio
- Re-architecting machine learning pipelines over time is part of the model life cycle
- The Art is to identify which workflow is the priority and have a robust strategy for re-architecting in-use pipeline

Carolyn Phillips is a Senior Staff Machine Learning Scientist at Wayfair.  She specializes in getting machine learning science research deployed at scale into engineering production systems.   Carolyn is passionate about building elegant, simple, but pragmatic solutions that make innovating easy.  With a PhD in Applied Physics and Scientific Computing from the University of Michigan, Carolyn began her career as a computational scientist at Argonne National Laboratory.   She is ready to wax poetically about the self-assembly of icosahedral quasicrystals if asked.

- What are text-to-image models and how do they work?

- How to use such models for business applications

- Live demos to showcase such technology

- Understand the need for context in Search today

- Various algorithms used in building embeddings

- The limitations of these models and how can we improve them

AI in healthcare and medicine has seen phenomenal advancements in the last few years and has enabled surgeons, scrub nurses, and pathologists to make well-informed decisions in clinical settings. As more AI-enabled technologies enter the healthcare ecosystem, AI teams face unique challenges like patient privacy concerns surrounding data collection, Health Insurance Portability and Accountability Act (HIPAA) compliance around data handling, and Food and Drugs (FDA) guidelines on product usability and safety. The requirement of explainability and interpretability also influences the choice of AI algorithms. We will dive into these critical challenges to better understand the ML product life and management in healthcare settings.

-  The product life-cycle in Healthcare AI must account for Health Insurance Portability and Accountability (HIPAA) compliance and Food and Drug (FDA) guidelines. 
- Given the severity of the impact of product failure (AI and the software that runs the AI), AI teams must take additional steps to eliminate software vulnerabilities and minimize model failures.
- AI teams must evaluate the tradeoff between explainability and performance by taking stakeholders into account.

Shubham Patil currently works as a Staff Deep-Learning Engineer and Researcher at Stryker, developing computer vision-based AI technologies that help surgeons, scrub nurses, and hospitals make informed decisions to improve health outcomes. Before joining Stryker, Shubham led AI deployment efforts at DawnLight Technologies for early patient fall detection alerts in critical care environments. He firmly believes in AI's potential for humanity's greater good and in serving our loved ones during their most vulnerable stage.
Shubham holds a Master's degree in Robotics from Carnegie Mellon University. He is passionate about developing enabling technologies using Computer Vision, Deep Learning, and Robotics.

- What are the common security challenges in federated learning and what are the established defense techniques?

- What are the current challenges when it comes to training federated learning

- How do we address challenges in managing federated learning and where should the research be focused?