2
$\begingroup$

I'm not really happy with the mind maps I've been able to find on Google, most of them are algorithm based. I want to make a good one that is problem/solution domain based. Do I have this right for my top level nodes? Here is what I have so far: https://i.sstatic.net/o7O2J.jpg

My questions/doubts about what I have so far are:

 

  • Is my starting point below generally correct? e.g. no high level subclass is missing, and everything presented as a subclass deserves to be here?

  • is Hybrid learning always just a combination of supervised and unsupervised? Or, are there real examples of other hybrid models (e.g. 'reinforcement' and 'supervised', etc.). I know theoretically we can combine any methods...I'm looking for what's real/applied/demonstrable today.

  • does Reinforcement learning belong at this high level, or is it actually a subset of one of the others (or one I've omitted)?

 

  1. Machine Learning

    1.1 Supervised (uses labelled data to train and validate)

    1.2 Unsupervised (uses unlabeled data, or ignores labels if they are present)

    1.3 Semi-supervised (uses partially labelled (mostly unlabeled) data)

    1.4 Hybrid (combines a supervised method and an unsupervised method)

    1.5 Reinforcement Learning (uses data from the environment)

 

Thank you!

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

Your picture is correct and can be used.

On the other hand, there is another perspective that can be used. Consider it as a hyper matrix with the following dimensions:

  1. Input objects. Scalar/Vector/tensor(picture)/sequence.
  2. Output objects. Scalar/Vector/tensor(picture)/sequence.
  3. Objective. Implicit/quality of representation (e.g. likelihood or KL divergence )/prediction error/immediate reward/long term reward.
  4. Prior knowledge. Not considered/implicit/explicit(Bayesian prior)
  5. Learning mechanisms. Random search/evolutionary algorithms/gradient descent/expectation minimization/variations Bayesian.

All these dimensions can be used for the categorisation of any algorithm. For example you can consider regression picture to scalar if you want to estimate age from photo.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you. My goal is to stay away from 'techniques and methods' for the first few nodes, and stay focused on problem domains. Do you have any criticism/feedback on the image knowing that it is focused on problem domains? I'm especially interested in fleshing out the 'regression' node with some problem domains. $\endgroup$
    – 3z33etm
    Commented Oct 29, 2017 at 14:57
  • $\begingroup$ For example, I miss time series models $\endgroup$
    – Karel Macek
    Commented Oct 29, 2017 at 15:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.