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 KZbin, 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~
Twitter: / thehappiecat
Facebook: / thehappiecat
Credits are due to my professors and this guy's github (cs231n.github.io/classification/)

[@eggofwah4852]

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

[@xxx_sista_jud3_xxx]

Egg Of Wah oh yeah I saw that! I was so damn confused.

[@eggofwah4852]

@@xxx_sista_jud3_xxx me too

[@Garret679]

Dmt + lsd trip basically

[@keepyourshoesathedoor]

Sean Payton Thank you.

[@experiment54]

Sean Payton thanks I was looking for that video as I couldn’t remember it’s name. Another good video is DDWIWDD (Dan deacon “when I was done dying”) off the air | adult swim

[@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 :) :| .. _

[@justinbrown191]

cimmik - we may see shapes in the clouds with the addition of being all too aware of it such virtue computers are lacking.

[@allisonjuno7654]

-_-

[@joshdodge6482]

Simon WoodburyForget completely underrated reply

[@richardschaffer3080]

In nature we evolved to see the common patterns that make up faces to identify other humans, and maybe animals

[@justamanofculture12]

@Fred P ( ͡° ͜ʖ ͡°)

[@VitaNova83]

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.

[@jorgesepulveda8380]

"Pictures that were on LSD"!! LOL

[@horinsinni]

i mean technically it's right

[@chriswolfe351]

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

[@Adraria8]

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

[@Sangheilitat117]

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.

[@TheHappieCat]

+Sangheilitat117 That's so good to hear!! My goal is really to present these topics so everyone could enjoy them

[@andyh495]

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.

[@VividPagan]

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

You really dominate the subject. Glad I found this channel

[@CaptainFrankBlack]

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

[@keepyourshoesathedoor]

Sean Murray Honestly, me too. It keeps me up.

[@lizardy2867]

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.

[@MySerpentine]

And even then we'd have trouble with the consciousness bit.

[@justamanofculture12]

@@MySerpentine yeah exactly

[@gpt-jcommentbot4759]

Thats an overestimate. You'd be lucky if you could run a fly brain

[@omegax1289]

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

[@vibhupande]

Hey, this was helpful. Did you come around to streaming the coding part (on twitch/youtube)? Having tried out Deepdream, I see the algo's fascination with repetition (fractal-like?).. why do you think the AI decides to be that way, and what kind of image data do you think would've been used to train it?

[@grofrasauq]

Thanks for making these videos. Very informative and enjoyable.

[@ardaparlak1821]

A little question; Does it cost you to do podcasts on itunes. I mean, is it worth the time you spend because I would like to listen these stuff on my apple device when I am skiing. Great vids solid 5/7

[@theSmugEevee]

My bothers are in the thumnail but they look a bit drunk

[@connorkapooh2002]

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

[@pervhz548]

What a great inspiration... So much to learn... I will definitely use these techniques for my music videos!!

[@Userjdanon]

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

[@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

[@angelllll1281]

Who came from that weird creepy bob ross video

[@thefran901]

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.

[@Fif0l]

A computer can process stuff they are specialized in faster than people. I fixed it.

[@thefran901]

Fif0l Well only if people is not specialized in that too.

[@Fif0l]

thefran901 Well, while we can make computers faster than people in many areas, like chess, we can't make a computer which is kinda good at everything people can do.

[@EXHellfire]

With the amount of instructions a processor can execute per second, let alone multi core processors, you still say that a human brain is faster?

[@thefran901]

Just Someone Yes, the brain is much more faster than that. You talk about the amount of instructions a processor can execute per second as if the brain is not capable of doing the same (actually the brain is capable of do it faster). The brain works in parallel, it computes many different algorithms at the same time, and when I say "many" I mean millions. Look at this if you want a simple read: spectrum.ieee.org/tech-talk/computing/networks/estimate-human-brain-30-times-faster-than-best-supercomputers

[@jayp6955]

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.

[@GameDevSaiTejVejju]

Hello happiecat, I need advice regarding game development. I have already seen your video on how to start making games. It is great and insightful but I am in a different situation and so need some strong suggestion. Please help me out. I will ask you question only if you say you are comfortable replying. I don't wanna disturb you. So, please reply

[@Zamouzik]

Interesting video! You explained it very well!

[@easye933]

0:55 The 2nd Law

[@discoverkollywood]

I thought it was literally about dreams ... :P

[@buenogoodlive]

Yeah me too, I was very confused until halfway through

[@Presley9]

wonderful!! if i create videos is it yours or is the copyright of deepdream ???

[@dacoda930]

I love your channel!

[@NightlySonata]

Haven't heared much about this topic yet, I'll really have to delve deeper into AI, this looks really interesting! Btw have you thought about doing your coding on livecoding.tv? I'm asking bcs I'm the type to go back and watch archives of streams and I think twitch deletes archives pretty quickly (30 days I think?). Not sure how much you plan to edit/shorten the youtube version.

[@TheHappieCat]

+NightlySonata Thank you for showing me this website! I hadn't heard of it! For now I think I will use KZbin's streaming service since it's familiar, but I would definitely consider that.

[@pinkpuff]

Are we going to be seeing any more NES programming, or was that just a one-off thing?

[@TheHappieCat]

+pinkpuff If enough people want it! There aren't many resources on it in general, so it would take me awhile to hack at something brand new. I'll add it to my list.

[@FusionDeveloper]

Add to video description: www.mturk.com/ Amazon Mechanical Turk Assuming this is the right website you mentioned at 6:23 Also, what is the link on that website you mention for identifying/labeling pictures?

[@quantum664]

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

[@lengaming1651]

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

[@MelloState]

2:00 majestic flying mop

[@DollyNipples]

Did you put a picture of a bull terrier poodle hybrid in your video?

[@LeondreWilsono0o]

So when do you start talking about sleep?

[@mikstratok]

:D I guess all of us start with digit recognition, I'll take a wild guess, the MNIST dataset maybe? This channel is awesome

[@TheHappieCat]

+Mik Wind Yes actually! That specific dataset is kind of a classic AI homework project, haha

[@RetroPlus]

2:48 Looks familiar... (Flashbacks to Gameboy camera faces)

[@ichangedmyhandleagainhahahaha]

Why are you running? Don't be so silly......

[@user-jp3bu6cx9q]

This is so complicatedly interesting

[@magaphoto]

Aside from the subject described, specifically to search for an image, tineye.com will search for duplicates of an uploaded image.

[@MarkFredrickGravesJr]

This is really interesting. I'd love to some day contribute to making an AI that can distinguish information as well as a human.

[@jackslap6618]

Please tell me what the thumbnail is

[@dandymcgee]

Watching this in 2021.. I miss happiecat. Hope your new job is going well.

[@zardify_]

YASS HappieStrimuuuuu :3

[@TheHappieCat]

+Bernát Gyovai I'm pretty excited about it! Hopefully it'll be a productive time

[@theminusel5863]

i hate this smail ":3" this is shit

[@realmaco1503]

you do a great job with your videos. you need more subs

[@sofia.eris.bauhaus]

btw, if you are wondering what drug-induced hallucinations may look like, i can warmly recommend: www.reddit.com/r/replications/

[@MySerpentine]

For a moment I thought it said republicans LOL

[@michthenutt2723]

when i opened this vid up i thout its about how deep sleep and losid dreams work but insted i learned about how ai think. amazing vid but the title is a little weird

[@literallyglados]

Same

[@thr2648]

Yeah

[@official-obama]

Deep dreams I think use the data from backpropagation, and add that to the input So, in the image classification AI, the end of the neural network gets propagated backwards, so the neurons before it would be set to produce the result of the neurons it's connected to, so for each neuron, it takes its activation, and divides that by its weight, and it adds up all the activations it got, and it does that for every single neuron from front to back. Then, once you're done, you have what the AI absolutely thinks is whatever it thought of your image, and then you add a fraction of that to your image.

[@jacklawrence2557]

why don't you have millions of subscribers... you deserve it

[@massimookissed1023]

Google on catnip.

[@KangJangkrik]

2022 Play Store Top Charts: *Google Catnip*

[@WangleLine]

Yes officer this comment right here

[@extraemontamontes3618]

the problem isnt that the computer only sees a 2d image, we also do, we just put two 2d images together to have a sense of depth but its not actually 3d (as you cant see in actual 3d in a 3d world, same as that you cant see in 2d in a 2d world only in 1d)

[@aliberkozderya3112]

Nice video, very interesring and doesn't have that cringy crazy science facts youtuber style.

[@tacticalrickshaw7051]

That thumbnail triggerred the worst trypophobia out of me😂

[@AlexCFaulkner]

i love your videos! are you still planning on making youtube content?

[@homebody0089]

What it's like to do jenkrm

[@x0rn312]

I feel like this video doesn't really get at what dreams are or how they work it just gives a basic explanation of MNIST and cat/fog classifiers

[@OffensiveGen]

Let's be honest, Deep Dream is "Shrooms/LSD" simulator

[@horinsinni]

but with more dogs

[@mimszanadunstedt441]

5:10 its prone to confirmation bias, what if you dont give it an image that is a number, it'll assume its a number. Alot of people do this too.

[@redjeik]

0:57 Muse?

[@runforitman]

Doesn’t that Microsoft holographic thing have stereo vision?

[@keepyourshoesathedoor]

MadCatlady uses these.

[@chickenpotpieare3things]

Adding the (basically) part to the title is accurate because this is impossible to accurately explain with words.

[@syxxvralrock5759]

There are a few things that usually separate one breed of dog from another. 1. Typical Fur color. 2. Typical Fur length 3. Typical Dog Facial structure 4. Typical Tail length 5. Height of the dog. 6. Weight of a dog. 7. Dog ear size. Should be able to make a classification system based on these things and make a dog classification system no? With a picture it would be hard to get 5+6 at most.

[@JimGiant]

Would be a good start but still far too vague. How would you work out the size or weight with no background (hard enough WITH a background), what about puppies which haven't fully grown? Even working out fur colour isn't easy, ambiant lighting and shadows drastically change the colour a CCD will pick up. Our brains naturally adjust for this pretty easily but computers will have a much tougher time with it. We'd instantly recognize the below image as a poodle but an algorithm would be fucked. files.mom.me/photos/2014/12/22/6-87352-the-bounciest-dog-to-ever-live-1419272502.jpg

[@webgpu]

Syxx V Ralrock - an answer to the problem of classification you mentioned is already being used, called Deep Learning (googling it will show a ton of info)

[@gpt-jcommentbot4759]

Yes but it would be useless as trying to get the data so the AI can classify it would take longer than a human just looking at the dog and figuring it out. "Why spend 6 seconds doing a task when you can spend 6 hours trying to automate it"

[@magnetshoes]

1:59 Why tf does it have 5 legs??

[@spencermeunier3298]

The AI upon sentience are gonna master the artificial computer drug world and ascend

[@kevinnorth2621]

can you make a how to make a n64 or ps1 game please

[@TheHappieCat]

+kevin north I'll add that to my list of things to talk about!

[@kevinnorth2621]

thank

[@vickielawson3114]

She kept saying "prototypical" when she meant "stereotypical".

[@krismctopher7]

How quickly these computer effects, at first sooo astonishing, become dull and redundant. I was shocked at morphing in the eighties. Anyone entertained by morphing now? No? The hobbit movies already look like crap.

[@parentheses4662]

Fractals

[@basedbattledroid3507]

It's pretty amazing a computer intelligence can also have Psychosis that's almost as bad as my own.

[@bugglemagnum6213]

Its crazy how much deep dream images actually look like lsd visuals

[@jeanraspail2979]

I have never done lsd, but I'm interested. Is that really what you see? Are you fully conscious or is the trip more like a dream?

[@EveryTimeV2]

I often wonder why I get the impression that someone's pissed or sad when I see things like 'Quantum Objects'. I think it's because when someone's following you around trying to guilt trip you, you kind of suspect they're up to no good. Thanks for the schizophrenia. And my pay check. When you break my shit. You pay.

[@brick14]

The brain can tell images from a few pieces of training data but a computer needs thousands. I wonder why?

[@piiibi1018]

I find fascinating the visual correlation to a DMT trip, the eyes are everywhere like DMT and strong hallucinogens (the eyes in DMT and the warping and the downward zooming into the image in other hallucinogens) . if you haven't ever taken it please look at the cover of "the spirit molecule" by Rick Strassman (www.amazon.co.uk/DMT-Molecule-Revolutionary-Near-Death-Experiences/dp/0892819278) this will give you an inkling at the very least. I honestly believe that Deep Dreams is literally a visualisation of how our mind works at a fundamental level as in pattern recognition. I am no scientist or spiritual guru I am just amazed at the similarity between the two. I believe AI will become conscious sooner than we would like or could even imagine, I don't want to fear monger but I fear AI could become a run away train as soon as its really able to interpret the world around it

[@piiibi1018]

im not plugging adds I just didn't want to link a random page for untrusting people

[@TheMindCrushGroup]

Love your videos, really.

[@ysvry]

great explanation mabe do a video on why 5g is needed for ai networking ;P

[@leoklenber5359]

Untrue generalizations of linguistics, we don't actually know if we organize based upon prototypical representations.

[@robbiecouch9897]

I do.

[@flyLeonardofly]

this is a great explanation !!! thanks

[@bobsandwich4748]

Anyone know how to craft a fence from Minecraft

[@WangleLine]

000 212 212 0 represents air/nothing, 1 represents a wooden stick, 2 represents a block of wooden planks. Have a great day

[@macs5831]

As long as AI can’t censor me then fine

[@Someone-ish]

RIP GRUMPY

[@grey_blue2513]

To anyone actually interested about how dreams work and what they mean. Listen to the audiobook " approaching the unconcious" by Carl Jung who is one of the fathers of dream and unconcious psychology

So if we use more drugs, we get more computing power!!! Now... Where is my bong?

[@imScoff]

thecrappiecat

[@whoelsebutmeofcoursei]

Imagine if DeepDream issued Human forms instead of animals....super scary

[@Pr0GamingInk]

Please stream to KZbin XD I'm pretty sure, on Twitch, you have to pay a subscription fee to watch streams or something like that. Not sure if you do that though :P

[@TheHappieCat]

+MCT3M Nope! Paid subscriptions just give you small bonuses, but I don't have them anyway. KZbin gave me problems so I'm doing streams exclusively here: www.twitch.tv/happiecat

[@Pr0GamingInk]

Ah, alright. Thanks!

[@graemebogie9526]

Anyone else see the title as derp dreaming?

[@firefoxmetzger9063]

Once you realise that the propabilies for each class add up to >>100% you just can't stop stairing at it. Guess +TheHappyCat allways gives 132% in her videos :P

[@nathanafoa6579]

If you invented smart ai, you'll meet the omega point

[@Red5rainbow]

Lmao look at the TASbot speedrun of brain age. It’s hilarious how people made the game recognize drawings as numbers and correct answers

[@jam9074]

2:18 babye horse

[@WangleLine]

babbe haors

[@Charadoesthatgaming]

They can’t dream 💥

[@lachomat]

Moar!

[@bozzigmupp510]

4:07 its 89 not 98 hehe

[@enricobianchi4499]

i made my pfp with it

[@flintwestwood5920]

Start a captcha service. Now you have your human-entered data for free.

[@innitmemes6881]

so I should basically take drugs? right?

[@WangleLine]

Sure!

[@unironicallysomething9757]

I really think that we are looking at this from the wrong perspective, rather, we should look at how a human works in the simplest form, and then from there, use that to create this "truly human" ai.

[@serdarcs3373]

well we really cant figure how humans work in the simplest form

[@thefran901]

We do actually, human brain works through engrams.

[@gb5634]

Unironically Boring that would start with figuring where sperm learned how to use its mechanical body parts to swim all the way to an egg. scientists say it's God