Inceptionism: Going Deeper into Neural Networks • Hacker News Board

Re: Inceptionism: Going Deeper into Neural Networks

#91

Post by IanCal » Thu, Jun 18, 2015, 2:49 PM UTC

abrichr wrote: post #66

IanCal wrote: post #15

The one generated after looking at completely random noise on the bottom row, second from the right: http://googleresearch.blogspot.co.uk/2015/06/inceptionism-go... Reminds me very heavily of The Starry Night https://www.google.com/culturalinstitute/asset-viewer/the-st... Lovely imagery. I never had much luck with generative networks. I did some work putting RBMs on a GPU partly because I'd seen Hinton talk showing s…

Neural networks are notoriously difficult to train due to the large number of hyper-parameters that need to be tuned. If your network never converged, it's possible your learning rate was too high, so it kept overshooting the minima of the loss function.

Quite possibly. The classification results were great, just wasn't good when trying to run things back through the network repeatedly. I did have issues that the learning rates reported in some of the original papers didn't match the ones in the released code.

Re: Inceptionism: Going Deeper into Neural Networks

#92

Post by drcode » Thu, Jun 18, 2015, 2:55 PM UTC

spot wrote: post #90

a really early version of this: http://draves.org/fuse/ published as open source in the early 90s. not NN but does have the same image matching/searching.

Fascinating link, but that was arguably different: In the case of the Google links, the neural network was built for other uses, and the "image fusion" is only a side effect... It is a sort of "proof" that some really interesting things are happening behind the scenes.

In the older approaches, the image fusion was the primary intent of the system. Still very cool, but much less impressive IMHO.

Re: Inceptionism: Going Deeper into Neural Networks

#93

Post by murbard2 » Thu, Jun 18, 2015, 2:59 PM UTC

Two remarks

1) Captain obvious says: the "tripiness" of these images is hardly coincidental, these networks are inspired by the visual cortex.

2) They had to put a prior on the low level pixels to get some sort of image out. This is because the system is trained as a discriminative classifier, and it never needed to learn this structure, since it was always present in the training set. This also means that the algorithm is going to be ignoring all sort of structures which are relevant to generation, but not relevant for discrimination, like the precise count and positioning of body parts for instance.

This makes for some cool nightmarish animals, but fully generative training could yield even more impressive results.

Re: Inceptionism: Going Deeper into Neural Networks

#94

Post by mortenjorck » Thu, Jun 18, 2015, 3:26 PM UTC

davesque wrote: post #25

This is one of the most astounding things I've ever seen. Some of these images look positively like art. And not just art, but good art.

This may sound ridiculous, but I think this has the potential to be a development as foundation-shaking as Modernism itself. There has been plenty of algorithmically-derived art over the past 30 years, but generative pieces inevitably look like math – they are interesting curiosities, sometimes quite beautiful, but they don’t challenge the mind like any of the major movements of the past 150 years.

This is different because, while still just math, it’s modeled on the processes of human perception. And when successfully executed, it plays on human perception in ways that were formerly the exclusive domain of humans – Chagall, DiChirico, Picasso – gifted with some sort of insight into that perception.

Future iterations of this kind of processing, with even higher-order symbol management could get really weird, really fast.

Re: Inceptionism: Going Deeper into Neural Networks

#95

Post by gradys » Thu, Jun 18, 2015, 3:28 PM UTC

Does anyone have a good sense of what exactly they mean here:

>Instead of exactly prescribing which feature we want the network to amplify, we can also let the network make that decision. In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at a different level of abstraction, so the complexity of features we generate depends on which layer we choose to enhance. For example, lower layers tend to produce strokes or simple ornament-like patterns, because those layers are sensitive to basic features such as edges and their orientations.

Specifically, what does "we then pick a layer and ask the network to enhance whatever it detected" mean?

I understand that different layers deal with features at different levels of abstraction and how that corresponds with the different kinds of hallucinations shown, but how does it actually work? You choose the output of one layer, but what does it mean to ask the network to enhance it?

Re: Inceptionism: Going Deeper into Neural Networks

#96

Post by pault » Thu, Jun 18, 2015, 3:29 PM UTC

davedx wrote: post #38

Worth reading the comments too. One from Vincent Vanhoucke: "This is the most fun we've had in the office in a while. We've even made some of those 'Inceptionistic' art pieces into giant posters. Beyond the eye candy, there is actually something deeply interesting in this line of work: neural networks have a bad reputation for being strange black boxes that that are opaque to inspection. I have never understood those…

I can't really put my finger on it, but this is the most exciting thing that I've seen on HN in years.

Re: Inceptionism: Going Deeper into Neural Networks

#97

Post by ebetica » Thu, Jun 18, 2015, 3:44 PM UTC

gradys wrote: post #95

Does anyone have a good sense of what exactly they mean here: >Instead of exactly prescribing which feature we want the network to amplify, we can also let the network make that decision. In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at…

My thought is that they basically run gradient descent on the image where the loss is the magnitude of one output plane in one of the layers of the neural network. Probably using gradient descent to push up the magnitude of one of the output plane layers or something like that.

Re: Inceptionism: Going Deeper into Neural Networks

#98

Post by romaniv » Thu, Jun 18, 2015, 3:58 PM UTC

fizixer wrote: post #71

Some comments seem to be appreciating (or getting disgusted by) the aesthetics but I think the "inceptionism" part should not be ignored: We're essentially peeking inside a very rudimentary form of consciousness: a consciousness that is very fragile, very dependent, very underdeveloped, and full of "genetic errors". Once you have a functioning deep learning neural network, you have the assembly language of consciousn…

We're essentially peeking inside a very rudimentary form of consciousness

Blatant sensationalism. There is absolutely nothing here that would suggest consciousness. If you have a mask for matching images, you can reverse that mask and imprint it as an image. What we're seeing here is a more complicated version of the same process. Heck, look more closely. Some of those "building" images have obvious chunks of pedestrians embedded, probably because the algorithm was trained on tourist photos.

Is it interesting? Yes, from algorithmic point of view. Cool as hell. However, this has nothing to do with consciousness.

If anything, some of those images are just a more elaborate version of a kaleidoscope. It's not like they run a network and got a drawing. They were looking for a particular result, did post processing, did pre-processing and tweaked the intermediate steps (by running them multiple times until the image looked interesting). Finally, we as viewers do our share of pattern matching, similar to how we see patterns in Rorschach inkblots. And there are captions that frame what we see and "guide" us to recognizing the right objects.

Re: Inceptionism: Going Deeper into Neural Networks

#99

Post by cafebeen » Thu, Jun 18, 2015, 4:04 PM UTC

sitkack wrote: post #67

This was recently posted to HN, http://tjake.github.io/blog/2013/02/18/resurgence-in-artific... Which mentions running the NN in reverse, quote By far the most interesting thing I’ve learned about Deep Belief Networks is their generative properties. Meaning you can look inside the ‘mind’ of a DBN and see what it’s imagining. Since a deep belief networks are two-way like restricted boltzmann machines you can make hidd…

It's worth pointing out that a naive bayes or k-nearest-neighbor classifier can similarly generate examples of valid inputs.

Re: Inceptionism: Going Deeper into Neural Networks

#100

Post by discardorama » Thu, Jun 18, 2015, 4:08 PM UTC

gradys wrote: post #95

Does anyone have a good sense of what exactly they mean here: >Instead of exactly prescribing which feature we want the network to amplify, we can also let the network make that decision. In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at…

My understanding: when you're doing standard gradient descent, you push the error down through the layers, modifying the weights at each layer. Now, in "normal" NN training you stop at the input layer; it makes no sense to tweak the error at the input layer, right?

But what if you did the following: flow the error down from the outputs to the layer you're interested in, but don't modify the weights of any of the layers above it; just modify the values of this layer in accordance with the error gradient.

Added later: I think we should wait till @akarpathy comes along and ELI5's it to us.