After a recent group activity in the Fuyao Program, I found myself thinking about the relationship between machine learning and human learning from a slightly different direction.
Machine learning has long borrowed ideas from human intelligence. But perhaps the loop can also run in reverse: after decades of building learning systems, we now have a more systematic language for thinking about learning itself.
Instead of only asking how to make machines learn more like humans, I am interested in another question: can the way we train and optimize machine learning systems help us become better human learners?
Why This Question Matters
Many of the core ideas in machine learning were originally inspired by human cognition and behavior: neural networks, reinforcement learning, attention, memory, and agents. Over time, these ideas evolved into practical tools for improving learning systems in a disciplined way.
We now routinely ask questions such as: what data should a model see, how quickly should it update, when should it specialize, what should it optimize, and how should it balance exploration with exploitation?
These were designed as machine learning questions, but they also sound surprisingly relevant to human growth. They point to a deeper possibility: machine learning may not be a literal description of how humans learn, but it can still offer useful abstractions for thinking about how we learn more intentionally.
Humans as Learning Systems
In some sense, each of us is a continuously trained learning system. We absorb inputs from our environment, update our internal model of the world, and carry those updates into future decisions.
Our learning rate also seems to change over time. When we are young, beliefs and habits can shift quickly. As experience accumulates, we become more stable, but also more resistant to change. Better priors help us navigate the world, yet they can also make us less willing to revise what we already think we know.
That is why continued learning may require deliberately placing ourselves in settings where our current assumptions are no longer sufficient. Growth often begins when an existing internal model stops explaining the world well enough.
There is also a human analogue to catastrophic forgetting. A model that is updated too aggressively on new data can lose capabilities it once had. People can experience something similar when short-term pressure, narrow specialization, or constant context switching erodes older skills, perspectives, or habits of thought that were once valuable. Continuing to learn is not only about updating quickly, but also about retaining what should not be lost.
Training Data, Curriculum, and Generalization
One of the most interesting features of human learning is that we partly choose our own training data. The people we spend time with, the work we pursue, the books we read, and the environments we enter all shape the distribution of experience we train on next.
That makes small choices unusually important. A person repeatedly exposed to one narrow set of problems, perspectives, and incentives can become highly capable in that local setting while struggling to generalize elsewhere. Diverse environments, interdisciplinary work, and unfamiliar conversations often function like productive distribution shifts: uncomfortable at first, but valuable for broadening what our internal model can represent.
Learning also depends on sequence, not just quantity. Machine learning systems often benefit from curriculum: building foundations before moving to more complex tasks. Human learning works similarly. Breadth becomes more valuable once there is enough depth to organize it.
This leads naturally to the tradeoff between specialization and generalization. Becoming an expert in one area is a legitimate and often powerful outcome. But breadth becomes meaningful when it is built on strong foundations rather than shallow exposure. The harder question is not whether to specialize or generalize, but when to do each.
Reward, Exploration, and Distillation
Machine learning also gives us a useful vocabulary for how learning is optimized. Early human learning often resembles supervised learning: we learn from examples, corrections, teachers, and demonstrations. Later, however, the most important decisions no longer come with labels. At that point, learning starts to look more like reinforcement learning, where judgment is formed through action, feedback, and revision.
That transition immediately raises a deeper question: what are we optimizing for? Salary, titles, publications, grades, and approval are useful signals, but they are still proxy metrics. Humans, like models, can become very effective at optimizing the wrong objective.
A similar tension appears in exploration versus exploitation. Relying only on what already works may trap us in a local optimum; exploring too broadly may prevent real depth. Strong learning, in both humans and machines, often requires a balance between the two.
Knowledge distillation offers another useful analogy. We do not need to learn everything from scratch. Teachers, mentors, books, colleagues, and now AI systems can all act as teacher models. The most valuable thing to learn from them is often not just the answer, but the process: how they frame a problem, where they focus attention, and how they respond when uncertainty appears.
What This Means for Human Learning
Taken together, these ideas suggest that better learning may be less about raw effort and more about better design. We can choose stronger training data, build better curricula, learn from better teacher models, revisit our reward functions, create room for both depth and exploration, and be more deliberate about retaining what matters as we adapt.
Machine learning does not tell us exactly how the human mind works, and it should not be taken too literally. But it does offer a disciplined way of asking good questions: what should we learn from, what should we learn first, how quickly should we update, what should we remember, and what objective is actually worth optimizing?
We have spent decades asking how to build better learning machines. Perhaps some of the same thinking can also help us become better learners ourselves.
Comments