A year ago I watched your MarI/O video and was exposed for the first time to neural networks. It was because of watching that video that I decided to take machine learning courses at my college. I am now pursuing a career in AI. Thanks so much!
If you're interested in learning more about neural networks, I highly recommend looking up "3blue1brown" on youtube and watching his videos about neural networks. It made my understanding of neural networks go way up!
I've been wanting to get into neural networking just recently so a refresher video of you demonstrating some of the cool things we can do is an awesome inspiration! :D
Wow, the trade off play idea was pure genius! Did you set it up so you could override the networks commands at any time, or have control alternate back and forth at a regular interval? I love the idea of taking control and guiding the network, like you're a master painter guiding the hand of your apprentice...... so sick!! Amazing work!
I just had it switching at slightly random time intervals, because I didn't trust my reaction time to be fast enough to correct the errors as quickly as I wanted. So I was just always pressing the correct buttons, whether or not I had control (even though it was only recording my button presses when I was in control).
This is very cool! If you are looking for something pretty simple to try and tweak this with, adding convolutional layers prior to your LSTM layers seems tailor-made for this type of application. Chase those gold cups :)
This video legitimately scooped a paper by a professor I was working with. "HG Dagger" uses the same idea of having a human take over control of an agent, and was published a year after this video
I am now enamored with the idea of having a tournament where everyone brings their own trained AI and then they compete. It'd be like an Amiibo tournament in SSB4, but for the Super Nintendo.
so last year. I did an essay for school on using neural networks for self-driving AI. and it failed miserably. both types of networks failed to go through a basic obstacle course. Spent a month's worth of time coding. Now I found that every major organization + Sethbling have done it and they make it look easy.
Oh man, I've worked with neural networks and classification with statistical information before. Seeing stuff like this make me feel like my experience is equivalent to handing a monkey a picture of a corvette and expecting him to assemble one.
Imagine training a RNN to fight enemies like you do for a whole JRPG. Then surprise, there's a secret dungeon where you enemy is a "clone" of your playstyle.
*google deepmind has alphazero. They now have AI for chess too and it beat Stockfish 20+ times (Stockfish is one of the strongest engine on chess, rated 3300+)
I remember I started watching this on Twitch and told my girlfriend that you do cool Mario hacking. When I turned it on you started to have a tantrum over the Bowser level, and I felt awkward and turned it off.
Amazing. As someone who just entered the world of Machine Learning, this video motivates me to become better and better at ML until I am also capable of creating NNs that can perform tasks like this.
Holy fuck, what next, fully functioning Hal 9000 and GLaDOS? Seth, you're just making the world 1 step closer to being overtaken by robots, LOL. Loving these projects, keep up the good work! You need a patreon
This is a subfield of machine learning called "imitation learning". You actually identified one of the major issues yourself, known as drift! The way you did it is one potential way of handling it. Though if you look up state of the art algorithms, there are more sophisticated fitness and randomness based methods for requesting human input. Overall, machine learning for this sort of domain is drifting closer and closer to reinforcement learning with no human input (often using self-play, forcing a computer to beat itself at the game over and over), but this is definitely a decent approach!
If this was implemented in games we could have interesting things. Imagine someone in Dark Souls being invaded by themselves, or having people train against themselves in competitive games for practice. Having bosses in some games be you or a friend of yours. Being able to summon yourself to battle alongside, well, yourself. The possibilities are endless. Mirror matches would be perfect.
Seth is trying to create a brain clone of himself so he can never really die.
+purplebootsgames I'd have this at my funeral as free merchandise for fans of my brain.
Welcome back, SethBling Reincarnation #4 here
SethDOS
One day you will be able to race Seth even though he's deader than a doornail
*THE RIDE NEVER ENDS*
A year ago I watched your MarI/O video and was exposed for the first time to neural networks. It was because of watching that video that I decided to take machine learning courses at my college. I am now pursuing a career in AI. Thanks so much!
The Modern Shoe ~ 👍
Similar story for me. I was focused on game design but I am switching courses around now to fit a machine learning path.
Rory Williams Wow, I am doing that too
"What is my purpose?"
"You try to mimick how I play in Mario Kart"
"...Oh my god"
mushroom taco lol
Poor lil' robot
That is why machines rose against mankind xD
mushroom taco That is a whole hack of a lot better than serving butter.
IM PICKLE RIIICK!!!!
If you're interested in learning more about neural networks, I highly recommend looking up "3blue1brown" on youtube and watching his videos about neural networks. It made my understanding of neural networks go way up!
Thanks for the recommendation!
LazerLord10 we
Same, I was just about to come down here and make my own comment suggesting the same thing.
For more interesting projects using neural networks I also recommend carykh
LazerLord10 thanks
I've been wanting to get into neural networking just recently so a refresher video of you demonstrating some of the cool things we can do is an awesome inspiration! :D
@ 3:42
"I recorded 15 hours of training data."
Otherwise known as playing the game. xD
Gotta give props to him for having the patience to play a relatively boring game for 15 h.
@@sontapaa11jokulainen94 Can be done in small increments, say 20min one day and 2h the next day maybe 30min the 3rd day etc
@@thisflyingpotato4227 yeah but still playing 15 h of some crappy (for todays standards) racing game is hella boring
You should probably get a github page going if you want to share source code? Much better interface than google docs :)
+orochimarujes
You should probably take a common sense lesson?
... and behind this is a bunch of command blocks!
Don't forget about the armorstands
kek he doesnt do minecraft anymore though
and behind _that_ is an original joke
Behind your comment is where you'll find a mountain of salt.
Minecraftfan4 hes still having a Redstone channel picture though...
*Dope*
*_Nope_*
666 likes original
Wow, the trade off play idea was pure genius! Did you set it up so you could override the networks commands at any time, or have control alternate back and forth at a regular interval? I love the idea of taking control and guiding the network, like you're a master painter guiding the hand of your apprentice...... so sick!! Amazing work!
I just had it switching at slightly random time intervals, because I didn't trust my reaction time to be fast enough to correct the errors as quickly as I wanted. So I was just always pressing the correct buttons, whether or not I had control (even though it was only recording my button presses when I was in control).
+SethBling I have no idea what either of you said but ok
They're talking about 4:27.
SethBling Nice! So, have you ever considered DefCon? Your content could be very interesting over there.
SethBling I thought that you had programmed it to relinquish inputting command when it detected input from an external controller
Good that I watched the 3blue1brown videos so I understood something!
SpektralJo thx
Same! I was like, "I actually know some of these words and the math behind it"
Oh also the sigmoid, weight, layers, etc.
O sht a bro
Same
That idea to switch between NN and yourself was pretty clever . I'm often impressed by your creativity man !
I'd be interested to see how it would react to a speed runner training it.
These neural network videos are so good. Please make more!
Currently learning neural networks, and this is gold. Thank you !
This is very cool! If you are looking for something pretty simple to try and tweak this with, adding convolutional layers prior to your LSTM layers seems tailor-made for this type of application. Chase those gold cups :)
Finally!!! More neural network videos! Keep it up SethBling, I love this stuff!
Miguel Ballard Agreed! Nothing I find more interesting really... you can apply them to an endless amount of problems!
Love it. Keep up the good work SethBling!
So glad you finally released the video on this!
this stuff is so incredibly cool! I love these videos!
I wonder how good it could get watching TAS runs
Combine the two systems... I think that would be interesting...
Thats what i was thinking too!
Easier said than done
You mean an LSTM(variant) "trained" via NEAT? That's a good idea that's sadly rarely done
Yes, why 200 neurons in 2 levels? Is that more than enough raw brain material that learning can take place, or can it be optimized further.
Really good, I'm personally working on a number recognizer so yeah, good video!
it was fun watching the videos in the background while I was working
This video legitimately scooped a paper by a professor I was working with. "HG Dagger" uses the same idea of having a human take over control of an agent, and was published a year after this video
I am now enamored with the idea of having a tournament where everyone brings their own trained AI and then they compete.
It'd be like an Amiibo tournament in SSB4, but for the Super Nintendo.
I have no idea what's going on, but I'm thoroughly entertained
so last year. I did an essay for school on using neural networks for self-driving AI. and it failed miserably. both types of networks failed to go through a basic obstacle course. Spent a month's worth of time coding.
Now I found that every major organization + Sethbling have done it and they make it look easy.
This was a great explanation! In case you want a more detailed version, 3Blue1Brown currently has a great series on that.
Oh man, I've worked with neural networks and classification with statistical information before. Seeing stuff like this make me feel like my experience is equivalent to handing a monkey a picture of a corvette and expecting him to assemble one.
Would be cool if we could get more interesting AI in games, using tech like this.
play against a bot that is trained by your own style of playing, that would be pretty awesome. :)
Imagine training a RNN to fight enemies like you do for a whole JRPG.
Then surprise, there's a secret dungeon where you enemy is a "clone" of your playstyle.
That would be beyond amazing.
Professor just used this video in our class, super cool stuff.
Google DeepMind has AlphaGo
Elon Musk has OpenAI
SethBling has MariFlow and MarI/O
Connor smart
*google deepmind has alphazero.
They now have AI for chess too and it beat Stockfish 20+ times (Stockfish is one of the strongest engine on chess, rated 3300+)
leela zero is an open source version around 2400~ you can play
*Elon Musk has left the chat*
Cipher
nailed the guess. i was staring ay the bottom left quadrant of gameplay from the start.
I remember I started watching this on Twitch and told my girlfriend that you do cool Mario hacking. When I turned it on you started to have a tantrum over the Bowser level, and I felt awkward and turned it off.
Lol I remember that time, Seth’s funny haha
So cool man! Great work!
Great work on the Skynet-Basics ! Keep on with your work
Thank you Seth - this is a brilliant introduction to NNs!
Jesus fucking christ dude, every video I see of yours is amazing. Love how smart you are my bro, keep up the good work.
Brilliant, as always. Thanks Seth.
Oh my, man, I love your neural network videos
Your projects are so cool!
Amazing. As someone who just entered the world of Machine Learning, this video motivates me to become better and better at ML until I am also capable of creating NNs that can perform tasks like this.
I got the notification for this video as I opened youtube.
GG
Lol you got me with Mr. Electric Mars.
This channel has had such a weird progression over the years. I love it
The man is back on TH-cam!
amazing!
i bet the best part was actually playing and recording your gameplay.
Holy fuck, what next, fully functioning Hal 9000 and GLaDOS? Seth, you're just making the world 1 step closer to being overtaken by robots, LOL. Loving these projects, keep up the good work! You need a patreon
This is amazing. Well done!
This is freaking awesome! Subbed!
As someone who just started learning Tensorflow. This is awesome!
I like how this channel went from Minecraft videos to neural networks videos
I love your content to bits.
We need a speedrunner to input data into this neural network. That would be interesting.
This sounds like a class project
Starting to find inspiration to get into this AI and Machine learning world. Awesome! Thanks for sharing!
So instead of learning from itself MariFlow learns from you? Neat
Self driving cars are also the future of video games
This is a subfield of machine learning called "imitation learning". You actually identified one of the major issues yourself, known as drift! The way you did it is one potential way of handling it. Though if you look up state of the art algorithms, there are more sophisticated fitness and randomness based methods for requesting human input.
Overall, machine learning for this sort of domain is drifting closer and closer to reinforcement learning with no human input (often using self-play, forcing a computer to beat itself at the game over and over), but this is definitely a decent approach!
Just like giant_neural_network pointed out, I had to comment about the geniality of the trade off play. Amazing idea!
0:57 I knew it. I put my finger to the screen when I saw him skid around like that.
That is some seriously impressive stuff.
As always this is so, so cool. lov u Seth. 👍
Being in calculus now your videos make sm more sense
making an AI to replace you? Doesn't try to be the best, just tries to be SethBling. real spoopy
Very cool, just subscribed!
Your idea is awesome.
And here I am just watching TH-cam videos. I ain't even mad. Good job. :D
yesss another seth video
I just love SethBling's neutral voice xD
90's : played mario for fun
10/20's : creating an algorithm to play mario for fun
Amazing! Keep up the great work.
We seriously need a neural network mario kart tournament.
Great video and great job!!!
This is actually what I was wandering
Interesting that this comes out in the middle of 3B1Bs Neral Network YT video series. Coincidence?
this is amazing, honestly.
now make a neural network that can complete dark souls
0:11 BEST joke ever, you got my like bro !
Definitely the best Minecraft youtuber out there
One word - *Astounding*
Sethbling the type of guy to play Mario Kart with calculus smh
The PMTTYD community used calculus to go fast 2yrs later lol
Very cool, makes me think of markov chains and cybernetic feedback.
man that was so mind boggling
Dude are you some kind of Mario supergenius?
If this was implemented in games we could have interesting things. Imagine someone in Dark Souls being invaded by themselves, or having people train against themselves in competitive games for practice. Having bosses in some games be you or a friend of yours. Being able to summon yourself to battle alongside, well, yourself. The possibilities are endless. Mirror matches would be perfect.
In feed forward the inputs are actual between 0 and 1 in most cases, to fit in with the sigmoids in the later layers :)
I remember when sethbling used to play Minecraft and I 100% prefer this sort of thing.
Welcome back sethbling here, today I will be curing cancer with Mario.
Also 10 seconds of footage and I could identify the human player.
YEAH I'M GOOD!
Hey again Seth. Looks like you have another mario neural network video.
You deserve a nobel.
This is quite possibly the coolest shit in the entire world.
Amazing! Thanks for sharing
I remember talking in the chat as "MariFlowAI"
Sethbling you will be the reason that the world reaches the singularity
This is the ultimate turing test!!
Train it with your data, then do a race against each other.
Tried the last MariIO. That was the best open source ML project I had tried on.