Pre-Read: TBD Topic

Date: May 14, 2026 Speaker: Vivek Natarajan (DeepMind)

Topic Overview

Topic TBD — check course site for updates. Speaker leads AI × Science × Medicine research at Google DeepMind, with focus on medical AI systems.

Key Concepts

Awaiting topic announcement.

Speaker

Vivek Natarajan

Affiliation: Google DeepMind, Research Lead for AI × Science × Medicine

Background: Leading researcher in applying transformers to healthcare. 18,300+ citations. Focus on making AI reliable in safety-critical medical domains.

Key contributions:

  • Med-PaLM — First AI to pass US Medical License Exam
  • Med-PaLM 2 — Expert-level scores on medical benchmarks
  • Medical AI validation frameworks — Patterns for deploying AI in high-stakes domains

Papers

1. Med-PaLM: Large Language Models Encode Medical Knowledge (Nature, 2023)

Contribution: First AI system to pass the US Medical Licensing Examination.

Key insight: Scaling + medical domain fine-tuning enables expert-level medical reasoning.

Result: 67.6% accuracy on USMLE (passing threshold: ~60%).

🔗 Nature paper

2. Med-PaLM 2: Towards Expert-Level Medical AI (Nature Medicine, 2025)

Contribution: Improved architecture and training, achieving expert-level scores.

Key insight: Instruction fine-tuning + medical reasoning chains improve performance.

Result: 86.5% accuracy on USMLE (expert doctors: ~87%).

🔗 Nature Medicine paper

Why It Matters for Autonomy

Aspect Relevance to Robotics/Embodied AI
Safety-critical deployment Medical AI is the gold standard for high-stakes domains — patterns transfer to autonomous systems
Validation patterns How to make transformers reliable in critical applications — testing, uncertainty quantification
Domain-specific fine-tuning Relevant to specialized autonomy applications (industrial, defense, space)
Human-AI collaboration Medical AI interfaces provide patterns for human-in-the-loop autonomy

Question Bank

Validation/Safety Questions

  1. What validation approaches from Med-PaLM translate to robotics/autonomy?
  2. How do you quantify uncertainty in medical diagnoses? Does that translate to action selection?
  3. What’s the process for detecting out-of-distribution inputs in medical settings?

Domain Adaptation Questions

  1. How much medical knowledge is in the base LLM vs added through fine-tuning?
  2. What’s the minimal dataset size for reliable domain adaptation?
  3. How do you handle edge cases where medical consensus is unclear?

Autonomy Transfer Questions

  1. Are there medical AI safety patterns that apply directly to autonomous vehicle decision-making?
  2. How do you balance model confidence with appropriate caution in high-stakes settings?
  3. What’s the role of human oversight in Med-PaLM deployment? How might that translate to autonomous systems?

Pre-Lecture Reading

Essential

Background

Cross-References

To be populated once topic is announced.

Prepared: 2026-04-04 • Topic TBD