All Convolution Animations Are Wrong (Neural Networks)

foreign animations you've seen up to this point are wrong and I don't just mean most of them I wasn't able to find a single fully correct animation seriously so I made my own but we'll get to mine first let's talk about what I've got against all the others you've probably seen these before because I think they're the most common animations out there from what I can tell they were first created by Vincent dumelin and they have undoubtedly helped many classes of students build up their first mental image of the convolution algorithm however there's one glaring problem these animations issue with 2D input and a 2d output but 2D convolution actually takes a 3D tensorous input and a 3D tensorous output and you've probably seen these 3D tensors if you've ever looked at a network architecture in a research paper like this one or this one or this one or you get the point now if you really want you can take the 2D animations and the 3D still shots from the research papers and combine them together in your mind and some people might be fine with this but I'm creating an introductory course on convolution and I don't want to give my students different pieces of the concept to put together in their head I just want to show them the end result and on top of that how am I supposed to explain more advanced concepts like depthwise separable convolution I don't want to try and paint a mental video with words and pictures I just want to show the video I looked around for Alternatives but they were all wrong in their own special ways this one turns the input and output units into 3D cubes rather than squares which is nice but it doesn't change much else this one does a nice job of showing the filter but everything is still 2D and it's not animated this one looks close right except that it's for a 3D convolution not a 2d convolution that's why the filter isn't the right size this one here is pretty close the filters aren't their true shape but I could probably live with that however I'd prefer a 3X3 filter so the viewers don't have to guess how the one by one generalizes to larger filters also you can see the interaction between the filter window and the output's appearance is a little awkward and finally the camera motion causing the world to flip around is a deal breaker I'm not trying to give anyone motion sickness here here's my favorite it's a relatively recent animation that shows a 3D input and actually makes a 3D output but without displaying the filters it's not obvious why the output is 3D and although the camera movement isn't as bad as the last one I'd still rather have it be stationary and those were the best that I could find so ultimately I decided that if I wanted this done right I'd have to do it myself that led me to spend way too long learning blender a 3D animation software which by the way the new geometry nodes feature is a amazing if I had tried to do these animations even a year ago it would have taken so much python code so shout out to the blender team and now for the final result here's one of my animations the input tensor on the left is its true 3D form with a padding of one in this case you can see each of the filters and their correct shape in the middle of the screen and the animation goes through each filter to form the full 3d output on the right the filter window is the common 3x3 size and finally the camera is stationary for the whole duration so nobody gets motion sickness I've posted this animation along with a few other variations on my github.io page with a link in the video description there are licensed under the MIT license so please share them and use them in your own projects I plan to build up this page to cover all the variations on Vincent's page and go even further to cover operations like depthwise convolution that absolutely require the third dimension I'll be posting new animations to GitHub as I release the corresponding videos here on YouTube so please subscribe and hit the Bell icon to be notified thanks for watching

Channel: Animated AI

Views: 54,871

Keywords:

Id: w4kNHKcBGzA

Length: 4min 53sec (293 seconds)

Published: Sat Oct 01 2022