How Deep Dreams (Basically) Work

Thought I'd do a topic I found pretty interesting: the algorithmic equivalent of LSD for AI.

Dev stream happening soon with more in-depth topics (like how to implement some of the stuff I talked about in this video!) I'm not sure if I'll stream on Twitch or Youtube, so let me know if you have a preference. I really don't know what I'm doing. Will announce when it's happening at least a few days ahead of time so follow me to know when it's happening~

[eggofwah]

Thank you for explaining the video called "deeply artificial trees"

[cimmik]

Isn't that the same way we humans "over interpret" what we see, so we see human faces whenever we can get away with it?

As soon as we see two dots and a line, we see a face, even though it looks nothing like a face :)

:|

..
_

[VitaNova]

The comparison between an over propagated neural network and drug usage is pretty fascinating! Or maybe just thinking of it as hallucinogenics tweaking the synapse weights.

[jorgesepulveda]

"Pictures that were on LSD"!! LOL

[Adraria]

I'm an aspiring game developer/programmer and I love this channel! Keep up the great vids :)

[Sangheilitat]

The amount of effort you put into these videos is awesome. I don't have any actual interest in pursuing game development or something like that, but these videos are just plain entertaining to watch.

[chriswolfe]

Do you have a drawing tablet, or are you using a mouse to do that? Either way, your drawing game is on point.

[andyh]

Good job! What I like about this video is that the bulk of it is an explanation of a specific basic image classification algorithm. I don't know how your other viewers feel but sometimes a ten minute video feels too short for some of the broader topics so specificity is good.

[MatthewJPrice]

what a trip back in time. for fun, i sent a snap of 2:07 to chat GPT and it quickly gave me this:


In the provided image collage, there are three dogs of different breeds. On the left, there's a small, slender dog with large, bat-like ears and a unique hair pattern that includes both hairless and furry patches; this appears to be a Chinese Crested Dog. In the center, there's a large, muscular dog with a sleek black coat and pointed ears standing in a noble pose; it looks like a Great Dane. Finally, on the right, there is a dog with a very distinctive corded coat, often seen in shades of gray or white; this breed is known as a Komondor.

What a time to be alive

You really dominate the subject. Glad I found this channel

[thefran]

It's not that computers are faster than people, well they are for a lot of things, but it's not because computers can process information faster than the human brain. The human brain is way faster than the best supercomputer. Computers are better for math than the human brain because they are programmed to do that, you could say that the human brain is a better hardware than the best processor out there, but the software we have is pretty limited. In computers we can program whatever software we want (generally speaking), we can't do that with our brains. While out brains are way faster than a calculator, they lack of the "calculator software". But they have the software to analyze images and find patterns in them.

[VividPagan]

Really great and understandable explanation for a complex topic, plus your voice is lovely to listen to and your doodles are fantastic. Thanks!

[CaptainFrankBlack]

I'm having deep intense dreams every night and wake up physically and mentally exhausted

[connorkapooh]

really cool to come back to this. i watched this when i was first learning about deep learning.

now im going to be an AI Engineer in under a month, pretty cool life goes by fast take it all in

[lizardy]

It’s not that AI is dumb, but more that humans have the specs beyond anything that could be made with standard computer parts. You’d be lucky if you could run a brain equivalent to that of a frog on current hardware.

[Userjdanon]

As a AI researcher, you are really doing a good job at explaining deep dream.

[lengaming]

1:41 holy shit a couple of dogs just barked outside

[quantum]

If we could get deepdream a training set based on sanskrit/greek letters and plants, and made it not quite violently overprocess images so much(keeping harmony, symmetry and the like), we could get close to an accurate trip simulation

[omegax]

I love learning about this kind of stuff. Great video!

[jayp]

At the 2015 Zillow/RETSLY Hackathon, one group built an app that pulled 2D images of a listed house, and tried to create a 3D virtual tour of the house, solely based on the 2D images. I am curious if they used the same algorithm as in 0:49, since I remember them saying they implemented an algorithm from one of Microsoft's papers.