3:39 - When you'll see it... (Row:2 Column:12)

@TwoMinutePapers

5 жыл бұрын

Whoa.

@sausage4mash

5 жыл бұрын

funkyprince is mad, do not remove your blindfold and look !

@jessehammer123

5 жыл бұрын

Column 2, Row 5 gets you some unholy combination of George Bush and Michael Bloomberg.

@DougSalad

5 жыл бұрын

Row 2, Column 10 is literally BLEEDING FROM THE EYES

@beskamir5977

5 жыл бұрын

Lovely, and I made the mistake of watching this during the night right before I'm about to go to sleep.

Is it possible to now use the inception scorer network to tweak the GAN network to maximize its inception score?

This is more enjoyable because you were in the past so thrilled at everything. What a time to be alive and all that.

Your channel is almost at 200k subscribers, keep the good work!

@martiddy

5 жыл бұрын

@@nagualdesign sorry, the autocorrect mess it up!

@martiddy

5 жыл бұрын

@@nagualdesign Actually, I'm not a fan of GamingGirl. But I have a friend who is called like that (is probably because of that).

@TwoMinutePapers

5 жыл бұрын

We are the most surprised by this. Thank you so much for all the love!

1: Inception Score as a measurement for the quality of an AI is the perfect name! 2: that sharkdog land squid at 5:20 is horrifying

thanks for describing "overfitting" for us. Helps us noobs out!

@mdp5337

5 жыл бұрын

I wish to recommend (in general, and it contains the notion) 'Algorithms to Live By' from Brian Christian and Tom Griffiths. Not in the specific context of NN based AI.

Inception Score (and its new version FID, Fréchet Inception Distance) measure differences of distribution of scores for 1// generated images, 2// real images-> bigger distance means less "real" generated images. Name "Inception Score" comes from Inception-v6 application (official Tensorflow version)that is used for this distribution comparison

Great episode as always!

Thanks for giving the hope, that there is a scope of improvement here. :)

Something like this would seem to be well suited for cosmetic surgeries. "Given that the network has been trained on celebrities, what is the closest face resembling my own?" or "How is my face different from that of a celebrity?"

Well, inception score doesn’t actually penalize overfitting. A network that just memorizes the dataset will get perfect inception score (and frechet distance as well). Afaik this is still an open problem.

Really enjoyed the video!

at 1:00 a lot of these images look good if you do not actually look but once you do they look really really freaky

2:55 L1 distance in pixel space? So manhatten distance, absolute value of the difference between the two images (pixel_i - pixel_i). Is that really a good distance formula for images?

@TwoMinutePapers

5 жыл бұрын

That is one way of measuring it. If I remember correctly, they also recorded the images that are similar in terms of neuron activations as well.

@joery8290

5 жыл бұрын

I've read in Goodfellow et al. 2016 That L1 norm measure in image recognition is usually preferred due to being able to resemble image polarity better. It's usually more empirical tho

0:11 - That's definitely Cory in The House

@tommycard4569

5 жыл бұрын

Hakim Mohamoud guess they can generate anime characters too

Thank u for this awesome video

Will there be identity generating NNs as well?

*TROOOLY EXCITING! WHAT A TIME TO BE ALIVE!!!*

0:15 Steven Wilson!

@jobigoud

5 жыл бұрын

All the images for the first 25 seconds are real people the network was trained on.

can you please do a video on google's BERT? thank you

Incredible

I'm not sure about the celebrity generator. I have a feeling it is memorising parts of faces like hair and noses and joining them together like a photo-fit. Instead of creating new hair and new noses if you see what I mean. It seems a bit too detailed to be entirely new. I would like to see it match the closest eyes and closest hair separately.

@hellfiresiayan

5 жыл бұрын

I agree. Many of the "bases" of the photos I can recognize as famous celebrity photos. For example, 3:33 top right corner is Kirsten Stewart, who had a very distinct haircut and skin tone at the time of the photo. The black guy at 3:35 has Dave Chappelle's very distinctive features. At 3:38 the fourth image on the first row uses the same frame as the Kirsten Stewart one. Finally at 3:38 as well, the 5th image in the 2nd column is 100% George Bush - it even has the Air Force One in the background. I suppose using a celeb as the base is expected since it can't be completely original, but some of the celebs have such unique features that it's hard to create new ones for the AI, I think.

Thanks for pointing out that Inception Score has limitations and is far from a perfect metric. It's important to keep information grounded to reality or we end up with unfounded hype.

I am looking for a movie of "Two Minute Paper" with an application that generates a 3D computer game and plays against this imagining (like as if we tried to solve imagined potential tough life situations) - can anyone help me with the link?... This 3D game looked like the old Doom, with orange brick walls.

4:34! 166!!!

0:40 Definitely a Hemsworth.

How easily can deepfakes and generated faces be combined?

4:55 - it generates cheeseburgers. surrender now you fools.

more papers to read

Don't forget mode collapse.

Are we sure these images are not copies

Am I the only one who is not that impressed by these generated images? I can clearly see which celebrities are blended in many of the cases.

4:45 "truly exciting, what a time to be alive" ...really?

Can NN make people more attractive? Can such a concept be reached to it?

WHAT A TIME TO BE ALIVE

I can see in the far future these will all be the faces of robots. Dun dunn dunnnnn! Spooky.

@sabofx

5 жыл бұрын

Funny how the text "Dun dunn dunnnnn" doesn't contain any melody whilst I'm pretty certain i know exactly what you mean. Makes me wonder who ever came up first with that tune?

Interesting what would our scientists work as if AI will ever have an IQ of over 200?

can you stop panning around in these image walls. makes me dizzy like crazy. thank you

1st like!

Long story short: can't solve a problem? Use a neural network.

generator.ai

Your use of English slang is over fitted :D

dont watch this shit high.