It accelerates many laborious and error-prone clinical tasks, including data abstraction, patient history summarization, and clinical trial matching.

(I make some brief remarks on this work in an interview at Meta: about.fb.com/news/2024/12...)

This last one is big, and has immediate real-world impact: we built an AI framework for processing cancer clinical text, centered on a custom embedding model trained from 8B tokens of real patient notes.

[3] "HopeLLM: A Cancer Specializing Framework for Multi-Data Abstraction from Patients' Entire Clinical Note Histories" by Lu, Mehdinia, et al. at the ML4H symposium.
ahli.cc/ml4h/schedule/

[2] "Empathic Coupling of Homeostatic States for Intrinsic Prosociality" by Yoshida and Man, at the IMOL workshop
neurips.cc/virtual/2024...

We implement a tiny version of empathy in multi-agent RL by coupling the internal drive states of homeostatic agents - and get prosocial behavior!

[1] "Need is All You Need: Homeostatic Neural Networks Adapt to Concept Shift" by Man, Damasio, and Neven, at the NeuroAI workshop
neurips.cc/virtual/2024...

We show that a homeostatic-like mechanism in neural networks helps them to adapt to concept shift.
Previous thread: x.com/therealkings...

👋

👋

👋