AI is starting to impress me.

Interdisciplinary science discussions. Also, if you are not sure where to place your thread, please post it here.

AI is starting to impress me.

Postby hyksos on June 18th, 2020, 2:33 am 

Artificial Intelligence has been making me roll me eyes for many years now. When AI is the topic of an article in popular science blogs and magazines, the article invariably contains hype, and sometimes the title of the article is completely misleading. (what some people call "clickbait") I'm not very impressed with AI. There was some rare glowing moments when I was inspired by Boston Dynamics videos. (But I could complain here about their robots being highly scripted, which I won't at this juncture).

The general clickbaitery and overhyping of AI has made me a very cynical person about this topic. So if something impresses me, it probably is actually impressive , and not just flashy.

There has been a sea change in the last few months, coming out of the GAN research. General Adversarial Networks. I'll show you what one of them did recently.

A GAN created by OpenAI was given the following image as input, and asked to complete the bottom of it.

Image


The GAN then produced this completion ,
Image

The GAN decided that the cat was wearing a black jacket, and even included some red cuffs sticking out of the sleeves. It gave him believable paws holding what looks like a pamphlet of stickers.

The cynical reader would look at these results and probably be able to guess why this isn't so interesting. The GAN network was likely just reproducing something it had been exposed to in the training set. In this case, this was not true. Because the corresponding image in the training set was this :

Image
The GAN actually made all that stuff up on its own, with the jacket and paws of corresponding fur.

This research is really different from what I'm accustomed to seeing. This is genuinely shocking, and I cannot begin to understand how it works. I'm feeling an emotion I haven't had in decades. It feels like I am impressed by this.

What the researchers did was that they had been working on very large models used to predict sequences of text -- what is called text completion , a sub-problem in NLP. They took those large networks and decided to feed them color images instead of text. The above were one of the results.

The reader of this article might be feeling as I am at this juncture, that what has been done here is so creative and unexpected that it doesn't seem real. There must be a hook about this research that makes it seem more palatable. The only thing that I found while reading was that they did not use full RGB triples for the images. Instead, each pixel is a 9 bit compressed representation, which the pub admits contains "some" information about color.

The next thing I will point out is the actual resolution of the images it was trained on and then later asked to complete. For demonstration I had blown them up for this article, but here is the actual output.

Image

more :

User avatar
hyksos
Active Member
 
Posts: 1823
Joined: 28 Nov 2014


Re: AI is starting to impress me.

Postby edy420 on June 27th, 2020, 6:46 am 

Kinda like seeing a toddler draw their fist picture of the house cat... only a little more advanced :P
User avatar
edy420
Active Member
 
Posts: 1347
Joined: 09 Jul 2010
Location: Fergusson st, Tokoroa, NZ



Return to Anything Science

Who is online

Users browsing this forum: No registered users and 19 guests