1
$\begingroup$

Can deep learning be a good way to learn a "High-quality" simple functions for images? For example, identical transformation, rotation, translation, even a linear mapping.

$\endgroup$

closed as unclear what you're asking by YiFan, NCh, Lee David Chung Lin, Lord Shark the Unknown, Leucippus Mar 25 at 5:22

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • $\begingroup$ What does "learn a 'high-quality' simple functions for images" even mean? Can you please clarify? $\endgroup$ – YiFan Mar 22 at 8:36
  • $\begingroup$ @YiFan sure, I think how to define "high-quality" accurately is a difficult problem. There is an easy way to do it: for example, we already have an identical transformation f. First, let's make a transformation f for cifar10, then test whether there is a huge difference between the classification results and result about the data without f for the same model. $\endgroup$ – Z.Huang Mar 23 at 9:26
0
$\begingroup$

The question is interesting , but vague , so may answer inherits the last property. In deep learning the models are highly nonlinear functions with lots of parameters . One consequence is that this models are very flexible - in the limit they can approximate anything, including any kind of linear mappings . However ,even if the data has a linear dependency , what you probably get after training such a model is a complicated non-linear function (but the quality of the predictions can be good ).

If you know that the data comes from a linear model, it is probably best to try to learn such a model using a more specific algorithm ( a random example: On Learning Rotations - Raman Arora ).

$\endgroup$
  • $\begingroup$ Thanks, I try to define a rule (about classification results) on this problem as shown in the comments section above. $\endgroup$ – Z.Huang Mar 23 at 9:30

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