Title

AWS re:Invent 2023 - Improve FMs with Amazon SageMaker human-in-the-loop capabilities (AIM334)

Summary

Romi Dutta, Amanda Lester, and Ketaki from Crikey AI presented on enhancing foundational models (FMs) with human-in-the-loop (HITL) using Amazon SageMaker Ground Truth.
Romi discussed the challenges in building, operationalizing, and using FMs, emphasizing the need for HITL for data labeling, model assessment, and fine-tuning.
Amanda demonstrated SageMaker Ground Truth's capabilities for creating demonstration data, preference ranking data, and captioning images and videos.
Ketaki shared Crikey AI's success story, detailing how they used SageMaker Ground Truth to train their AI animation foundation model, saving time and costs.
The session highlighted the importance of HITL in the AI development process, despite the advent of FMs trained on large, unlabeled datasets.
SageMaker Ground Truth offers comprehensive HITL capabilities, including model evaluation, data collection for model customization, and access to expert workforces for data annotation.
The presenters showcased the practical applications of these tools in generating high-quality, labeled datasets for various AI tasks.

FMs, while powerful, still require human judgment for tasks like data collection, evaluation, and fine-tuning, which is where HITL services become crucial.
SageMaker Ground Truth provides a scalable solution for HITL tasks, offering both automated and human evaluation workflows, as well as access to specialized workforces.
The HITL process is iterative and requires continuous evaluation and fine-tuning to ensure models are accurate and free from biases or toxic content.
Crikey AI's use case demonstrates the practical benefits of using SageMaker Ground Truth, highlighting significant savings in time and resources for startups and enterprises.
The session underscores the evolving nature of AI development, where tools like SageMaker Ground Truth are essential for managing the complexities of training and deploying robust AI models.