Convolutional Neural Networks



Katie! Hey Ben okay, so you told me what we're talking about. Today I've been thinking and thinking and thinking about. Puns and jokes stories and I I've come up with a couple of them, but honestly they're just too convoluted. To work. See what you did there. WAS ON. Narrow network. is a little weak, but. As best. I can do today, so you know that's what you get. Like the result of a Max pooling operation on all of your weeks off puns I did a week sauce punts oh. Man. I'm sorry. Hey, don't know is Max, pool, a person wait okay, we'll get to this in a second. You're listening to linear, digressions. Max Pooling is a layer in Kabul on that, but I'm probably getting ahead of myself a little bit, aren't I okay, because it sounds like the name of a person. Max Pooling Hi. My name is Max. Pool Yeah Max Power. That's a person. No, but it's a joke on the simpsons where homer he's on the spot has the makeup of fake name for some reason. And he looks at some kind of. Like a power tool or something? He's like Max. Power I forget what yeah this would be. I, guess you. You kind of generally went to avoid things pooling in the sense of just. Puddles all over the place. De Yeah I. Yeah I can't think of a use of that well. I mean like maybe those those kids inflatable pools like. But that would be like you get Max Pool, but not. Maybe Max Bullying. Okay well anyway, We? So, what does that visit actually? It's not about. Exploring really bad puns and it is not is about. Conversational neural nets which is a particular type of neural network architecture for image, recognition and classification. That kind of thing do you know why it's called conversational own that because conversational meetings complex difficult to follow? Idea actually did they like to know? I would like to know I mean. That's why I asked the question. Because it seems kind of strange to be like there is technique a technique. Be An difficult technique. Well, so it's almost it's keeping you out. It's kind of interesting well, okay, we're getting ahead of ourselves a little bit. Can you give me give me a couple of minutes of preamble, and then we'll get conversational part. Okay, yes, sounds good. So. This is a neural net architecture that is particularly good like we said for emiss- recognition and classification, and we've talked about neural nets a lot in other episodes, so I won't go a ton into the details here. If you've never heard of neural nets before just a the quickest introduction I can think of so the idea of neural network is. It's a algorithm architecture where you have these things called while they're sometimes called neurons actually. And, they take a number of inputs. Come originally, in the case of let's say image recognition there the pixels in the image those pixels go into basically a mathematical function in each one of the neurons. And that mathematical function takes all of the inputs multiplies them by a bunch of different weights, ads them together, and puts them through some kind of activation function and. If all of the inputs times, all of the weights sums up to larger than some number then the neuron has a one as an. And if it doesn't go above that number that it has zero output. Grossly oversimplified right, and so you have. So you have these layers than of many. Neurons in each of the layers and each layer feeds the outputs from each layer, the inputs into the next one, and then after one or two or three or four end of these. At the end you have a prediction about in this case. What kind of object was in the picture and so? Convoluted neural nets or a particular architecture of neural networks that are best Ramos recognition. And they have special types of layers in them. That's what the presence of certain types of CON-. They're called. Constitutional layers in neck is one of the signatures of. The fact that it's a constitutional neural net. Okay so you've got all these layers that feed into each other, and if some of those layers are conversational errors. Then you've gotTa Constitutional, Neural Net? What is it constitutional? There yeah a conversational air one the the reason that you have all these layers of neurons is because when effectively allows a neural nets to do in this context. Is it learn more and more complex features? And what a conversational layer does is kind of interesting. A picture is going to be depending on how you look at it. Either a two D or three D object the two D. objects, if you think of it as just a bunch of pixels that are arranged in Two d grid. It's a three D object. If you WANNA, think of each of those pixels as having not just a one zero values that it can take, it can actually have red, green and blue, and then a mixture of those three colors. If it's a color image, so if you wanted to have all of the information represented into a picture, you could do it in this three dimensional cube of information if that makes sense. So imagine you got your picture and I think the best way to explain. A conversational layer is imagine that you're picture is a bunch of bricks, and they're assembled into a wall, so I know that bricks are rectangular and pictures are usually square. You'RE GONNA. have to stretch the metaphor a little bit. Let's imagine. This big wall made up of bricks in each one of the bricks can be different color, and so that's how you're painting your picture, right? We can call them bristles. Sorry? Okay, so got your. You got your bricks. And they're making your. They're making your picture now. Imagine that you have. A piece of poster board. And, so you? Can you have this piece of poster board and you can slide it. You, know you're kind of on a a crane. Let's say that's a forklift that's up against the wall and you can slide this piece of poster board over the wall. And so at any given point, this piece of poster board is over a bunch of Bristles, and it can like kind of look at what is going on under its under its shadow. Okay. This is like a sentient poster board yeah, now. invert the metaphor slightly so now instead of it being a piece of poster board that's blocking out part of the picture. Instead imagine that by default. The whole picture is blocked out, and instead you kind of sliding around a window, and so as this window is sliding over the wall. It's scanning. All of your bristles and it's. Doing something with them, right? I'm basically building up a metaphor for what a conversational areas so. You're sending the pixels into your neural net. And then there's what's called a filter or colonel, and that's your piece of poster board so usually it's something like three pixels by pixels or five pixels by five pixels towards that yeah. It's not particularly big, and it's this cutout this going to be sliding around and it's going to be looking at all of the different little sub regions of the picture and trying to figure out what's going on in

Coming up next