What is machine learning?

Machine learning is a way of developing computer systems by allowing them to learn patterns from examples. Instead of writing a separate instruction for every situation, developers define a learning process and supply data. The resulting model—the learned component—can then apply those patterns to information it has not encountered before.

Consider recognizing handwritten digits. People write the same number in different sizes, styles and orientations. Trying to describe every possible variation with explicit rules would be difficult. A learning system can instead examine many examples and adjust its internal settings to distinguish useful visual patterns.

Those settings are often called parameters: numbers that influence how the model transforms information into an answer. Training is the process that adjusts them. Inference is the later use of the trained model to produce a result. A model does not necessarily change its parameters every time somebody uses it.

There are several ways to organize learning. Supervised learning uses examples paired with target answers, such as photographs labeled by object type. Unsupervised learning looks for structure without supplied target answers or category labels. Self-supervised learning creates targets from the data itself, as when a language model predicts a missing or next piece of text. Reinforcement learning uses feedback on actions and their consequences to improve future choices. These approaches can be combined within a larger system.

Learning is not the same as memorizing. A model may perform impressively on familiar examples and fail when the same task appears in a slightly different form. Generalization is the ability to apply learned patterns usefully to new cases. Testing therefore needs examples that were not used to teach the model, and that resemble the situations in which it will operate.

The examples also influence what can be learned. A digit recognizer trained only on clear black numbers against white backgrounds may struggle with blurred photographs. More examples are not automatically sufficient if they repeat the same narrow conditions or contain incorrect answers.

Studying care pathways using a center’s historical records makes those human choices visible. Which patients are represented? What was known at the decision point? Does the outcome describe what happened or what should have happened? A learning algorithm can find patterns in the supplied examples, including patterns created by local capacity constraints. It cannot decide whether those are the patterns a future clinical service should follow. Defining the question and the reference outcome comes before choosing the learning method.

If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.

AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy