Title

AWS re:Invent 2022 - Ultra-low latency machine learning at Amazon Ads (ADM301)

Summary

Varun Kamlakaran introduced the session, highlighting Amazon Ads as a key strategic partner for AWS and the focus on ultra-low latency machine learning.
Shenghua Bao discussed the Sponsored Products team's use of deep learning models for product understanding and trend identification, processing tens of billions of impressions daily.
The team faced challenges with memory limitations and network consumption, which they addressed with a scalable hybrid approach for online feature serving, binary serialization, and traffic optimization.
They also tackled real-time shopping trend detection using AWS services like Kinesis and Lambda, and developed strategies to handle bot traffic.
Pooja discussed the Ads Moderation team's use of AWS services to improve the moderation process, focusing on four key services: Amazon OpenSearch, DynamoDB, SageMaker, and Recognition.
She detailed the evolution and optimization of these services for ads moderation, including cost reduction, latency improvement, and precision in decision-making.
The session concluded with insights on leveraging AWS service updates and features to maintain state-of-the-art systems.

Amazon Ads relies heavily on machine learning to provide context-aware sponsored product recommendations and to moderate ads, ensuring customer trust and advertiser success.
The Sponsored Products team's approach to handling large-scale online feature serving and real-time trend detection demonstrates the importance of scalable and efficient systems in a high-traffic environment.
The Ads Moderation team's use of AWS services illustrates the continuous need for optimization and adaptation to maintain cost-effectiveness and performance.
The evolution of AWS services, such as the transition from Elasticsearch to Amazon OpenSearch and the adoption of Inferentia instances for SageMaker, shows AWS's commitment to providing improved solutions for machine learning applications.
The talks underscore the importance of staying current with AWS service updates and features to leverage the latest advancements for improved performance and cost savings.
The strategies shared by the speakers, such as the scalable hybrid approach for feature serving and the configuration of archival strategies for DynamoDB, provide valuable insights for organizations looking to optimize their use of AWS services for machine learning and data processing tasks.