Ctrl+K

Define learn

Define learn#

Test#

What does it mean to learn (see also Learning)?

As a subgoal, we’d like to more precisely define what the article Reinforcement learning calls the three “paradigms” of machine learning (in the first paragraph):

See also Machine learning - Approaches for another attempt to relate the three paradigms.

Value#

It’s a popular word. If “some machines” can learn (see Machine Learning) should we learn to think like them, or they like us? Although this blog typically lets Wikipedia define terms, we will deviate from that default slightly in this case to allow for precision and “decompress” the definition a bit.

Cost#

Learn from data#

If you take “experience” as a synonym for “data” then this definition of machine learning (from Machine Learning) implies “learn from data” is insufficient to define machine learning (lacking a definition of T and P):

A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.

We’ll continue to use TPE to decide what fits in at least this definition of machine learning.

Learn for credit#

If you take “performance” as a synonym for “credit” then this phrase is also insufficient to describe machine learning (according to the TPE definition), since it lacks a definition of E. However, most people would argue this phrase implies some kind of test; few people would bother to argue that a “read” or “learn” task without acceptance criteria is equivalent to getting paid to do nothing.

Why#

The word Why is used (confusingly) in both Causal inference and to query the value behind an action. Does a toaster warm up bread because it’s hot or because that makes it delicious? It’s likely this is because we have functions to estimate value in our heads, and to backpropagate on those value estimation functions we reuse the same word we would use for any mental function backpropagation.

Life goals#

Should you be thinking about your life goals with every step you take? For an extended discussion on how to learn for credit, see:

Learn by example#

Generally speaking, the phrase “learn by example” is associated with Supervised learning. An example defines “truth” (the P) and more examples (E) should improve performance, so an algorithm that will “learn by example” should generally be considered a machine learning example by the TPE definition. You can consider this a special case of learning for credit, if you assume that the truth is good. This is sensible and Normative; see Truth - Folk Beliefs).

Supervised learning is related to Reinforcement learning, though it’s hard to say that either is a special case of the other. In SL the value signal is defined in a static way (e.g. through a function that operates on static data relative to the output) or a semi-static way (semi-supervised learning). Both methods get down to a single value in the end, but the SL objective function is much more detailed than the RL objective function (at the cost of being based on a static environment i.e. the risk of concept drift). RL is explicitly defined to include “state” in the model’s function, but many supervised learning algorithms include state (e.g. signal processing filters and language models).

Does four years of college without an application in mind make sense? Yes, if you accept one of multiple value signals. More static (supervised learning) algorithms might try to e.g. mimic your professor’s mental networks exactly (learn to think the way they think, including everything they say), or get good grades (perhaps based on your own understanding of a topic), or at least get the final degree (avoid failing grades). A more flexible reinforcement learning based model would test what you learned from the professor on some real environment (e.g. a lab-based course).

Unsupervised learning#

I’d argue that Unsupervised learning is a form of machine learning where the performance measure P (see above) is how well the algorithm discovers “features” that are interpretable by humans i.e. that we can fit into our natural languages. We might say that a neural network used for a vision application learns what an “edge” is, but the definition of “edge” it is using is complex, more complex than we can describe in plain language. See also:

Statistics#

The article Machine learning references this comparison between machine learning and statistics:

Statistics draws population inferences from a sample, while machine learning finds generalizable predictive patterns.

See Review. This author would argue that this is the difference between causal inference and statistical inference, making causal inference equivalent to machine learning (and defined more precisely, in general). You’ll see the same argument all over Statistical Rethinking. Statistics is about correlation, and correlation does not imply causation.

Compression#

Can an idea that appears in a dream teach you something, or data generated by a simulation? Consider the analogy of learning to compression. This analogy asks us to think of the model (with its hyperparameters) as a form of compression of some real-world process. If you accept this analogy (or accept it as a synonym), then you can still learn from a simulation if the simulation provides some valuable alternative approximation of the real-world process you can use to generate data (E). To use a simulation of reality, we assume we want to match a real-world process i.e. reality i.e. the Truth valuable, defining P.

This is not to say that machine learning is a compression of the data generated by the real-world process; the data may be completely insufficient to describe the process (or excessive). Said another way, machine learning (or a neural network) is the approximation (compression) of a function.

Another article exploring this topic:

Contents