Google’s Multimodal Med-PaLM with Vivek Natarajan and Tao Tu
Topics covered
Popular Clips
Episode Highlights
Collaborative Efforts
The development of Med-PaLM M at Google is a testament to the power of collaborative efforts. highlights the cross-functional nature of the project, emphasizing the contributions from various experts, including clinicians, engineers, and legal advisors 1. This interdisciplinary approach ensures that all aspects of the project are meticulously handled, from data set preparation to legal compliance 2.
We are not training these models from scratch, but rather we are building on top of the authors of Palm and Palm E.
---
adds that the team includes doctors who are also Google employees, providing critical expertise and feedback throughout the development process 1.
Interdisciplinary Input
The interdisciplinary input in Med-PaLM M's development is crucial for its success. explains that the Pathways framework, a large-scale ML orchestration system, allows for efficient training of multimodal models on thousands of accelerator chips 3. This system is optimized for AI workloads, making it possible to handle the complex computations required for medical AI.
Pathways was Jeff Dean's dream and vision, and many others have worked on it.
---
shares his motivation for working on this project, emphasizing the mutual benefits between AI and human intelligence and the potential for AI to accelerate scientific discovery 4.
Collaborative Frameworks
Google's Pathways framework plays a pivotal role in the collaborative development of Med-PaLM M. describes it as a system that allows for the training of giant multitask, multimodal models on thousands of accelerator chips, optimizing both computations and communications 3. This framework is essential for managing the vast amount of data and compute required for such a project.
It's a large-scale ML orchestration system that allows you to train giant multitask, multimodal models on thousands of accelerator chips.
---
The flexibility of the Pathways framework, combined with the robust infrastructure at Google, enables the efficient scaling of AI models, ensuring that they perform well across different tasks and datasets 5.