Hacker News new | past | comments | ask | show | jobs | submit login
Comparing the hallucinations of four similarly trained neural nets [video] (youtube.com)
54 points by bemmu on July 16, 2015 | hide | past | favorite | 13 comments



Could someone give a bit more background on what hallucination mean in context of neural nets?

I'm not sure what the goal or methods of achieving it are?


http://googleresearch.blogspot.com/2015/06/inceptionism-goin...

If you add a few constraints, you can run a neural network "in reverse". Crank up the "dog" output (which is now an input) and it will generate an image that would be recognized as a dog. You can also feed in images like normal, then modify the output and run it backwards to make the image more "doggy". And you're not limited to the first and last layers; you can enhance the values of different layers in the middle to bring out smaller or larger features.

Here's a live stream that might help you visualize it better: http://www.twitch.tv/317070


Okay that inceptionism article is REALLY cool. This was my favourite image: http://4.bp.blogspot.com/-FPDgxlc-WPU/VYIV1bK50HI/AAAAAAAAAl...

It's a single image of a "cloud" as has been interpreted by a neural network trained to recognize animals... so it interpreted many of the cloud shapes as different animals!


Don't miss the gallery link, it's got a high-res version of that photo and lots more https://photos.google.com/share/AF1QipPX0SCl7OzWilt9LnuQliat...


As I understood it, you ask the net what it sees in the image and then make it adjust the image so that it sees more of that thing.

So if a cloud looks slightly like a dog, it will make it even more doggy.


What's funny is that it all still relies on our ability to see dogs in things that aren't dogs.


Dogs in particular seem to drop out of the nets I've played with very regularly.

I did find the lizard setting eventually though...


I depends on what the networks have been trained to recognise - if they have been trained to recognise images of dogs then they are more likely to detect dogs in clouds. See also https://news.ycombinator.com/item?id=9818077 .

And see also https://news.ycombinator.com/item?id=9736598 for more background to the topic.


FYI, this topic is also known as hashtag #deepdream.


I was under the impression that it was referencing the similarities with drug induced human hallucinations.


It is, psychedelic induced hallucinations (as opposed to hallucinations as a result of delerium) are produced in much the same way as deepdream. And the results are strikingly similar.


was thinking about this, all the deepdream images we've been seeing look really similar because they all use the same network, would be interesting to use custom trained networks more variation


Well, really they look similar because nobody bothers to use anything but the default training set it comes with - which happens to contain tons of dogs.




Consider applying for YC's W25 batch! Applications are open till Nov 12.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: