Physics Net | Ifu Aniemeka | OpenAI Scholars Demo Day 2018

so hey everyone thanks for showing up tonight Thursday night Phyllis my name is equal a new Mecca and I'm a software engineer from Chicago and today I'll be talking to you about physics net I couldn't come

up with a clever name for it so we're still gonna call it and as you can see it's it concerns training a model to be able to predict the motion of objects in a environment so physics nerds prepare yourselves alright so from a pretty early age human beings are able to do some pretty complex things right so within roughly a year of being born you should have been able to throw things you hopefully were also able to track falling objects with your eyes and within a year most children can pick up objects and unfortunately put them directly in their mouths temper this little super smart doesn't do that kind of thing so without being told at any point you sort of gain beginning intuition that say for example if you throw a ball into the air it'll follow a roughly parabolic trajectory you also know that if you throw this ball at a wall or whatever I'm sports not my thing you know that it will it'll bounce off of that wall so the purpose of my project was to try to train a model to figure out how objects move in space without explicitly being given concepts like force or momentum just to like to get it from the same way that weekend in which is via observation of course all of this is in service of eventually constructing a robot army because you know they need to understand that you have to aim where the target will be and not where it is so to create the environments I used JavaScript physics engine called meta jeaious so you can kind of see from the way the circles are moving this doesn't perfectly mimic the real world all of these collisions are perfectly elastic there is no spin so there's no angular momentum there's no air surface friction so effectively if you keep the recording going the balls just travel forever so this is a fairly elementary thing to do for a human being right so on the far left you'll see circles at time let's say like T equals zero right that's position 1 the frame in the middle that would be time like T equals 1 the outward circles are in a different position in that last frame you can see some vectors that are being drawn between the initial position of the circles and their final position right so we kind of do this subconsciously and because we're able to do this so easily we can guess the next position of each of these circles you see these the circles outlined in red or obviously like the final positions in the third frame and this is basically what I asked my model to do I give it two frames with circles in obviously different positions and then I would have it guess what's the third frame where will the circle be at the next time step so the model wasn't given images act exactly like these are rectangles and you know a white space it's actually given samples like this they're somewhat pixelated the actual samples are 28 by 28 pixels so we can blow them up they look like this but that is actually enough information for the model to be able to figure out what it needs to figure out so this is the architecture of my model it's a convolutional auto encoder so as you can see at the top with the encoder I what I did was I took two images and I essentially you stacked them and I fed them through several convolution layers and at the end of that the decoder at the bottom the decoder is the portion of the network that actually essentially like undoes the convolutions and reconstructs the image so at the top I give it two to the initial images and then at the bottom it constructs that third image I'm gonna give you a sort of a quick rundown of what the model produced during training but to start with it's not really producing much of anything right it's just a gray square after about 2000 epochs it figures out that it should generate walls but as you can see the walls are not the right color they're white and also there's nothing inside of the box after about 4000 epochs it actually has figured out the color of the walls they're gray which is awesome but still nothing then it starts producing smudges after about 6000 epochs you can see like in the middle there just purplish something going on and in the lower left corner there's something greenish that's kind of what we're looking for over time it gets more accurate this is pretty close I constructed in animation just to give you a better idea of what it predicts so on the left is the target and on the right is what the network predicts I constructed the animation on the right by taking the two input images that I gave the network and then concatenating it's prediction on the end of that so you can see it's pretty good with velocity position specifically right after each time step but the colors are a little off right like the circles are flickering so there's some imperfections here that could be fixed but I was pretty happy to see that you know it's it has an idea of where the circle should be so next s right the point is for a model that generalizes to how the world works using not too many samples right so one of the things I want to do is essentially want to add complications to this environment one would be to add barriers or like walls inside of the environment to see that it understands that great wall means bounce off right another thing I could do would be to

change the environment shape which is another sort of generalization around bouncing off walls adding friction there's there was no friction in the original environment fixing the issue around the colors in the frames keeping them stable and another thing I don't

really mention here is adjusting the framerate so how I chose the framerate was a somewhat arbitrary but I don't know if that was a noble thing essentially I wanted to have a high enough framerate that if a circle bounced off a wall it wouldn't be such that for example you know it's like a little ways again from the wall at time step one and then like pretty far away from the wall at time step two I wanted enough frames per second that I would I would capture it getting pretty close to the wall and you know eventually bouncing off so I'd like to see you know what happens when I when I adjust the frame brakes either higher or lower just like it's kind of see what happens so yeah there's a lot I could do with this and I'm really excited to you know experiment some more if you want to find me I'm on twitter at ing Becca you can go to my personal site at life is algorithm calm and you can find me on github Larissa told me not to say Terminator she called it the T word so I see I didn't Larissa I didn't say it I just put it I didn't well I did its we're fine we're safe everything's good so yeah thanks for coming give any questions you can feel free to ask them now and so I'm not sure how to rephrase your question but you're sort of theorizing that maybe additional computation would help deal with like the color issue we should talk later okay let's talk about this I don't yeah I'll show you the code and you can fix it for me any other questions you ever because if we never trained it on c3 oh um I'm not sure that it does I are you commenting on I'm sorry so the question is about how would the model dealt with the blurring like the blurry pixels Oh ml magic eye yeah I didn't do anything like special to get that to happen it just like oh that's your question okay so yeah you're the question is around like how over time it figured out that the balls are not like smeared and actually sort of brought them together into like a single link central location um not sure we know how to answer your questions see no well yeah we'll talk about it later yes yeah thank you for coming don't be - when we're back in play