2
$\begingroup$

I have a presentation about ML vs DL. A basic one at that. I had a project about CNN and Robotic manipulator, which I finished, so let's say I know a thing or two about AI.

So, before I start ML vs DL, I want to start with classification, what is it etc. I am planning to say that, "ML and DL, in basic, are classification programs."

I mean at the end of the day, what ML and DL does is (after the training I mean) classify an information that is given to it, such as is it apple or orange, did the voice say this or that or is this data cluster fits here or to there.

Am I missing something? Or again would it be correct to ML and DL classification programs?

$\endgroup$
1
$\begingroup$

DL is a subset of ML, so I'll comment on ML particularly. You'd be wrong because classification, although an important one, is only one of the tasks we perform. Some other problems addressed by ML methods can be listed as follows:

  • Regression (many classification algortihms also do regression)
  • Clustering (k-means, hierarchical clustering etc.)
  • Dimensionality reduction (autoencoders, pca, t-sne etc.)
  • ...

It's even possible that classification problems can be though of as regression of the probability that the given sample belongs to a certain class or not.

$\endgroup$
1
  • $\begingroup$ Oh I see. I forgot that, you can use these algorithms to do tasks using prior examples, such as creating human faces or dlss in games. Thank you for the answer. $\endgroup$ Feb 17 at 11:11
2
$\begingroup$

This is too coarse. A classification algorithm maps an input to a discrete category prediction, or a predicted probability distribution over discrete categories. I can think of at least three other common uses of deep learning models:

  1. Regression models map inputs to continuous predictions, like stock prices or temperature.
  2. Generative models define a probability density over potential inputs, and can be used to generate e.g. simulated images. Examples include Generative Adversarial Networks (GANs) and Variational AutoEncoders(VAEs).
  3. Embedding models map inputs to vectors so that inputs for similar vectors are near to each other. Examples include BERT with fine-tuning, where BERT is pre-trained on a huge corpus of natural language and then "frozen" so it can be used to provide vector representations for subsequent language tasks.

Also, I question the framing of Deep Learning vs Machine Learning, since Deep Learning methods are a subset of Machine Learning methods.

$\endgroup$
2
  • $\begingroup$ For the ML vs DL part, it is more of a question to be answered such as "is there difference and if there is, what is it. For the other part, I have noticed the mistake in the assumption. $\endgroup$ Feb 17 at 11:14
  • 1
    $\begingroup$ A classification algorithm is even coarser than @jkpate implies. It maps predictors to a class membership which represents a forced choice as in "all or nothing". Classification bypasses the more logical continuous probability prediction. See fharrell.com/post/classification. $\endgroup$ Feb 17 at 11:56

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.