4
$\begingroup$

I had a brief experience with machine learning by using a clustering algorithm, i also read the basic ideas and calculations of a simple classification algorithm. Now, i would read more about "machine learning" and I found many similar definitions like the following:

"Machine learning is the science of getting computers to act without being explicitly programmed..."

"Machine learning is a type of artificial intelligence (AI) that allows software applications to become more accurate in predicting outcomes without being explicitly programmed... "

My understanding is that the ability to learn and adapting deduction (output) without reprogrammation is the main idea, and based on my personal understanding, this "adaptation" is only possible with "supervised algorithm" with a new training which can permit a change/adaptation/improvement on the output model with the same program and source code.

So based on my understanding again, this "adaptation" and "learning" definition doesn't fit unsupervised machine learning algorithms, since the model with all calculations is fixed and implemented! Any change will need an update on source code!

Therefore, I would have corrections to my misinterpretation, and more clarifications to have a better understanding of "machine learning" and unsupervised/supervised learning relation.

$\endgroup$
  • $\begingroup$ I guess this is a very good question. Maybe it's because you first initialize the parameters randomly most of the time then you update them, learn, to reach to desired outcome. $\endgroup$ – Vaalizaadeh Feb 19 '18 at 21:03
1
$\begingroup$

You said:

and based on my personal understanding, this "adaptation" is only possible with "supervised algorithm"

However, it is not true. Let's consider a clustering method as an unsupervised algorithm like K-means. Each iteration in K-means algorithm is a basis for the next iteration of the K-means up to the convergence of the algorithm. Hence, it means we are learning the structure of the data in each iteration up to reaching a specific target value. Although it is unsupervised, we are learning the structure of the data in each iteration. Indeed, it is a kind of (unsupervised) learning at all.

$\endgroup$
1
$\begingroup$

Supervised learning refers to learning a concept from examples. Those examples typically require work from a human, which is usually "expensive". In fact a huge part of practical ML-work is to come up with smart ways of obtaining training data (think of reCaptcha's for instance).

Unsupervised learning is learning without labeled data, that is all there is to it. Clustering is one example, PCA is another and autoencoding is the newest hottest thing (imho). In some sense unsupervised learning try to compress the data (by finding subspaces of high dimensional representations along the lines of which the data is ordered, or manifolds): You are trying to formulate concepts that describe the data on a higher level, this involves finding smarter descriptions.

It might be somewhat of a leap from K-means, but consider this: I could easily describe a photograph to you using a set of abstractions (fi: I see a dog with a bone -> 23 bytes). This would give you a picture of the scene without the need of me sending you the full dataset (fi: 1080 * 768 points ~ 10^6 bytes). Conceptually this would be somewhat similar to: 4 examples of cluster one and one example of cluster 4, instead of sending 5 complete instances of, say, 10.000 features.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.