Table 1.

Comparison of supervised, few-shot, and zero-shot machine learning.

Learning approach	Description a	Examples and analogies
Supervised learning	Model is trained (“fine-tuned”) on a labeled dataset, mapping input data to the correct output by minimizing error between predicted and actual labels.	ML model to distinguish between dogs from cats trained on 1,000 pictures of dogs and 1,000 pictures of cats, each labeled accordingly ML model to identify documented GOC discussions trained on 5,000 EHR notes, of which 300 are labeled as containing GOC discussions
Few-shot learning	Pre-trained model is fine-tuned to, or prompted with, a small number of labeled examples, relying on the model’s pre-training to apply previously learned patterns to a new task. Examples need not cover full range of variability or noise expected in the target task, as the model leverages prior learning to generalize effectively.	ML model with generalized pre-training fine-tuned on just 6 pictures of dogs and cats to distinguish between them A child learning to distinguish airplanes from helicopters after being shown 5 examples from a picture book LLM prompted to identify documented GOC discussions after being given a small number of examples
Zero-shot prompting	Model is prompted to make predictions without any task-specific examples, relying entirely on the model’s pre-training to generalize to the new task.	Vision-language model with generalized pre-training asked to distinguish between dogs and cats, with no fine-tuning or task-specific examples Resident physician asked to characterize an unusual chest x-ray finding LLM prompted to identify documented GOC discussions by definition alone, with no provided examples

Abbreviations: ML, machine learning; LLM, large language model; GOC, goals of care.

^aIn machine learning, pre-training refers to the initial process of training a model on a large, general dataset to learn patterns, representations and features that are broadly applicable across tasks. Pre-training is often performed by the model’s creators. Whereas, fine-tuning refers to subsequent adjustment of a pre-trained model on a smaller, task-specific dataset, often performed by those adapting the model for a particular use.