Crowdly

x, y (set of samples and set of labels), with y being discrete, qualitative

x, y (set of samples and set of labels), with y being continuous, quantitative

When training a clustering model, what kind of target function (f : x ↦ y) are we learning?

f(x) := Probability(x) — Density estimation case

f(x) := class_of(x) — Classification case

f(x) := expected_outcome_of(x) — Regression case

Relate each training set prototype with the kind of ML problem it can enable.

Training set in the form:

x (observation) -> discrete y (target)

Training set in the form:

x (observation) -> continuous y (target)

Training set in the form:

x (observation)

Relate each definition with the description of available data and feedback.

We have observations and labels, labels are in the same format as the ones our system will have to predict. Our learning process is driven by the difference between each expected and predicted outputs.

We have observations, our system makes predictions, and receives occasional rewards.

We only have observations. But we usually have a lot of them.

To learn something, what do you need to do?

Remember all the experiences you faced, with as much details as possible.

Forget experiences' details to keep only general ideas.

It this possible to perform machine learning with a simple database system (i.e. storing exact information and be only able to search using exact indexes)?

Yes it is.

No, this is not sufficient.

Let's say you have a database with many entries in this form:

situation -> action to perform if we face this situation

What would you need to turn this system into some ML-enabled system that can handle unseen situations?

Be able to add more entries to the system.

Be able to measure the similarity between an unseen and a previously-seen situation.

Nothing, this system is already ML-enabled.

Relate each term with its definition.

Gap between the best possible solution in our search space and the optimal solution, caused by approximation error.

Sensibility of the optimal solution found (in our search space) to the training set.

Small errors in the data which causes inexact search space exploration.