Some Thoughts on Practice and Learning by Humans and Machines
Learning as practice, learning as iteration, in clinics and in code
The joke, told in various ways, its that it’s called the practice of medicine because it’s just that: practice. As with many enduring jokes, there is an uncomfortable insight that makes it stick. Of course, medical practitioners know that medicine is not a scrimmage. The uncomfortable insight is that practice is a way of learning, and who wants to confront the fact that their doctor, however experienced, may not yet know everything they need to know? Yes, we want our experts and professionals to keep learning. But the counterpart of ongoing learning is that practitioner knowledge is inevitably incomplete. Who wants to think about the fact that their doctor will learn tomorrow what they needed to know today to best treat me? Or, even more uncomfortable, maybe I’ll be the patient that the doctor learns from?
I teach a class on data and clinical decision-making at Yale. When developing the syllabus, learning turned out to be a surprisingly important and revealing lens for several key topics in the course, including artificial intelligence. Defining what artificial intelligence is, to me, an exercise in futility (if not an unintentional exercise in human self-flattery: we are, one way or another, talking about our intelligence). Also, humanity has spent innumerable hours trying to define what is meant by intelligence, and no consensus has emerged. It’s not like adding computers into the mix will make arriving at a consensus definition any easier.
I also would argue that a lot of what we think of as artificial intelligence has overlap with cybernetics. The Wikipedia definition of cybernetics is useful (if not cumbersome):
Cybernetics is the transdisciplinary study of circular causal processes such as feedback and recursion, where the effects of a system's actions (its outputs) return as inputs to that system, influencing subsequent action.1
In this definition is a sense of learning: outputs being shaped as inputs to “influence subsequent action.”
Norbert Wiener2, the key pioneer of cybernetics, wrote a seminal book on the topic, and the title of that book serves as a much more relatable definition:
Cybernetics, or, control and communication in the animal and the machine3
Indeed, a sampling of chapter titles highlights the relevance of cybernetics to what we call artificial intelligence:
Chapter 5: Computing Machines and the Nervous System
Chapter 8: Information, Language, and Society
Chapter 9: On Learning and Self-Reproducing Machines
However, further digressing on artificial intelligence versus cybernetics will only reinforce the fact that intelligence is a slippery concept that, to me, is almost impossible to satisfactorily nail down.
In contrast to intelligence, learning is a much more concrete concept. Haselgrove has a definition of learning that I like:
[learning is] a relatively permanent change in behaviour as a consequence of experience4
Haselgrove’s book covers learning in the animal kingdom, but it has relevance to artificial intelligence. Indeed, the less flashy term “machine learning” is, to me, a much better framing for what we (essentially colloquially) call artificial intelligence. The key insight of machine learning is that computing machines can be programmed to learn rules from data. Traditional computer programming executes a logical or mathematical rule written by a human for interpreting data. In machine learning, a computer is given input-output pairs of data from which the computer learns an optimized rule that maps the inputs to the outputs5. The ability to learn from data is a critical insight. However, what turned out to be revolutionary is not so much that the rules are “intelligent” (whatever that means) but rather that, with sufficient data, these machine-learned rules significantly outperform human-designed rules for task such as image classification (cat versus dog versus car and so on) and natural language processing.
One of the ways that computers learn is that they go through a training process that, in essence, allows a program to change its behavior in light of the performance during training. Learning by practice, if you will. After training, the machine-learned model is tested and validated, meaning that the model is fine-tuned and verified that it can perform at a required level of accuracy before it is used in the real world.
In professions such as medicine, learning starts with classroom education or more modern variations on the theme. The “real” learning subsequently happens through observation, practice, and then by doing the actual thing in the real world. As the saying goes in medicine: “see one, do one, teach one.” As with jokes about practicing medicine, this aphorism reflects an uncomfortable insight.
The origins of the phrase are revealing. It originated in the late 19th century at Johns Hopkins by the surgeon William Stewart Halsted. Dr. Halsted is credited with developing the first formalized training model for surgeons, and “see one, do one, teach one” was key to this model. Quoting from Kotsis and Chung6:
Halsted’s model of “see one, do one, teach one” is based on acquiring increasing amounts of responsibility that culminated in near-independence. Halsted was not only interested in developing a system to train surgeons, but also in creating teachers and role models. A formal training program was the only way to ensure that surgical advancements would be passed on efficiently and effectively.
Whether the task is surgical or not, we know that you need to see the thing and do the thing in the real world to (however imperfectly) learn the thing.
It also is worth noting that teaching, in this model, is part of learning. Far from the lazy invective that “those who can’t do, teach,” those who can do are in a unique position to teach and, indeed, have a unique obligation to teach. Reading into the Halsted model, and bringing in my own (non-surgical) experience, few things reveal one’s humbling gaps in knowledge and expertise as having to teach.
I like the use of the word “independence” by Kotsis and Chung. Independence suggests mastery but not perfection. There will be situations that an experienced practitioner encounters that perhaps no one has encountered before. Or that this individual practitioner has encountered only a few times before, which hardly counts as thorough “practice” for mastering that scenario. But independence means that the practitioner will navigate the uncertainty and unknown and, importantly, use the experience to update their approach. That is, they will continue to improve and learn.
Humans can learn on relative small quantities of data using relative little energy, especially compared to computing machines. The explosion of data centers and attendant explosion in energy consumption to support large language models is a testament to this fact. It’s also worth noting that most machine learning models can’t really learn in real-time, something humans are very good at. For example, as most people have by now probably encountered, large language models have a knowledge cutoff date. The cutoff date represents the state of knowledge of, in essence, the internet at the time of large language model training. Even when apps such as ChatGPT search the Internet for knowledge generated after the knowledge cutoff date, the retrieved knowledge doesn’t get incorporated into model itself. It isn’t learning as it goes.
I’ve always found it fascinating that one of the most intuitive (to me, at least) approaches to machine learning turns out to be among the least useful in medicine. This approach is called reinforcement learning, defined as:
A computer program interact[ing] with a dynamic environment in which it must perform a certain goal (such as driving a vehicle or playing a game against an opponent). As it navigates its problem space, the program is provided feedback that's analogous to rewards, which it tries to maximise.7
Reinforcement learning is best known for leaning to play games of quite high complexity by using the outcomes of individual games to essentially learn the rules of the game over the course of an inhumanly large number of iterations. It is, roughly speaking, learning in real-time.
A vivid illustration of this learning can be seen in a DeepMind agent learning to play the block-breaker game Atari Breakout8. Reinforcement learning can achieve superhuman performance in classic games such as chess and go. It also can achieve superhuman performance in open-ended video games such a StarCraft. To do this, though, it requires an incomprehensibly large amount of data. For example, the equivalent of thousands of years of gameplay (practice, really) were required to achieve human performance of the video game Dota9.
This is why reinforcement learning has a difficult time in medicine. Even if you could formulate aspects of medicine in a game-type manner with game-type outcomes that serve as feedback for learning, it’s not clear if you could generate decades / centuries / millennia of gameplay. In contrast, as a rule of thumb, the time for a human to achieve specialized expertise is on the order of 10,000 hours, which is a little more than one year. Yes, those 10,000 hours are spread out over a a few to several years: humans eat and sleep, after all. Even if you work 80-100 hours per week during training, not every hour is weighted the same for productive learning (wrestling with electronic medical records is a reality, for example). That 10,000 hours of specialized professional training sits on top of 1-2 decades of classroom education. As I’ll sometimes joke, graduating from my MD/PhD studies represented graduating from the 24th grade.
This isn’t to say that human learning is better than machine learning. Or, to say that machine learning can never eclipse human learning. It’s more to say how different they are.
Cybernetics. (2025). In Wikipedia. https://en.wikipedia.org/w/index.php?title=Cybernetics&oldid=1300921342
Norbert Wiener. (2025). In Wikipedia. https://en.wikipedia.org/w/index.php?title=Norbert_Wiener&oldid=1301152822
Wiener, N. (1961). Cybernetics: Or control and communication in the animal and the machine (2nd rev. ed.). MIT Press. (Original work published 1948)
Haselgrove, M. (2016). Learning: A Very Short Introduction. Oxford University Press. https://doi.org/10.1093/actrade/9780199688364.001.0001
Kotsis, S. V., & Chung, K. C. (2013). Application of the “See One, Do One, Teach One” Concept in Surgical Training: Plastic and Reconstructive Surgery, 131(5), 1194–1201. https://doi.org/10.1097/PRS.0b013e318287a0b3
Machine learning. (2025). In Wikipedia. https://en.wikipedia.org/w/index.php?title=Machine_learning&oldid=1301513312
Google DeepMind’s Deep Q-learning playing Atari Breakout! (2015).
Reinforcement learning. (2025). In Wikipedia. https://en.wikipedia.org/w/index.php?title=Reinforcement_learning&oldid=1301026985