5
$\begingroup$

From both my own exploration of Google Deep Dream using Dreamify for IOS, and from Google Image results on the topic. I've come to 3 conclusions about the networks understanding of images that seem too common to be coincidences.

I would like to know if these are recognised as consequences of the approach, or otherwise to what degree thease properties are imagined while trying to understand the ambiguous high frequency images the Deep Dream algorithm often produces?


  1. 3 dimensional understanding of real and dreamt objects, including rotation, occlusion (close things covering distant things), reflections etc...

  2. Understanding of how objects group and are arranged allowing deep dream to create believable landscapes, settlements and narrative and physical interactions, including between people.

  3. Objects (including people) being thematically/stylistically designed to fit a scene.


For example:

https://photos.google.com/share/AF1QipNgj7mRqnzW5G2D7ty3Sqb_ZITjnBQZA23RH71691sSNVGBvUTVmdSCzb_3b360xQ?key=SHpzeTVRbHk4SHFhUmMtRXVPRGxuMGRnd2xrM1dn

$\endgroup$
  • 3
    $\begingroup$ It's not really clear to me how to define the claim "There is an understanding of (...)"; I'm not sure how we'd define what it means for a deep learning network to have an understanding of something. I realize this doesn't really answer your question. $\endgroup$ – D.W. Sep 16 '15 at 22:34
  • $\begingroup$ Thanks, clarified in the question. By understanding I mean, being taken into account, and therefore being part of the calculations. $\endgroup$ – alan2here Sep 16 '15 at 22:42
  • 2
    $\begingroup$ It seems that question is a specific example of the general problem of Deep neural nets as black boxes. Looking at the result of the output of the network we can think of "concepts" the network understand BUT when looking at the network all you can see is large number of function compositions. $\endgroup$ – Ankur Sep 17 '15 at 6:56
  • 1
    $\begingroup$ are you running experiments? plz drop by Computer Science Chat to give more detail if possible $\endgroup$ – vzn Sep 17 '15 at 17:08
3
$\begingroup$

"Understanding" is a human concept ie a term used for human psychology. It may be used with machine learning systems informally. However, being flexible in interpretation of your question & not strictly literal, there is of course an open scientific question/ research program of how these systems actually "work". In ML, one "trains" systems and the systems exhibit the desired "behavior" but some of this relates to an old psychology theory of "behaviorism" where only the observed behavior is seen to be relevant and attributing particular mental states is seen to be something of a problematic/nonobjective ("anthropomorphic") projection by the experimenter.

With that caveat, the current scientific understanding of (deep) neural networks, with lots of related research and active/ongoing, is roughly that they build/ use feature detection. This overlaps with neurobiology theory and knowledge of the visual cortex in mammals. The exact features that are evolved (via "emergence") in the network are somewhat unique to each network, but general trends have been observed also, namely orientation columns (eg esp in "lower levels"). These types of features are also known to arise with sparse coding techniques/ mathematics.

There is a hierarchy of features. Higher level features are built out of lower level features inside the network hierarchy.

So the answer is roughly "yes", the network develops "spatial/ visual concepts" in the form of features. The networks recognize structural patterns in common across many images of real-world objects. These types of patterns include different/ various orientations, placements, relative spatial relationships, and sometimes more abstract "styles" of images found eg in paintings/even architecture etc.

$\endgroup$

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.