Gerald Filimonov

My trials and tribulations in creating some lapping plates. Inspired by videos by Tom Lipton over at oxtools So a while back I got inquisitive and started wondering how the first flat surface was made. I mean sure you can compare something to a surface plate, but how was THAT surface plate made and verified to be flat without something else flatter to compare it to? I mean it's like a chicken and an egg problem right? So down the rabbit hole I went and stumbled upon the Whitworth 3 plate method, Tom Lipton and well... here I am. =)

Written in C++ using OpenGL for 3D rendering. This is a demo of an app I wrote to learn more about how neural networks do feed-forward and back-propagation operations. Feed-forward, I had a pretty good handle on but knew nothing about how to do back-propagation and wanted to learn the math behind the process. All of the AI code and the 3D renderer are written from scratch. The AI code runs in a separate thread on the CPU and is not GPU accelerated at all. Speed is currently throttled at 1 image per second. Full speed is 20ish images per second. You can fly around in space in real time using the mouse and WSAD keys for movement. The rotating view for most of the video was just a hard coded animation just for visual aesthetics. Dataset is the MNIST numbers dataset that can be found online. The dataset contains 70,000 images handwritten digits with each image being 28 pixels high and 28 pixels wide along with a file containing the labels that correspond to each number. Training is done in real-time as you're flying around but in this video the neural network image I loaded had trained for 20 epochs. Project was heavily inspired by: Denis Dmitriev: https://www.youtube.com/watch?v=3JQ3hYko51Y and Marijn van Vliet: https://www.youtube.com/watch?v=Tsvxx-GGlTg I will eventually upload the code to GitHub, but I'd like to take some time to fix a few issues first and clean things up a bit. Just some nit picky things. =) Update: Ok, so I'm probably not going to get around to fixing up the code so I've just released it as is. Keep in mind, I was learning as I was writing this, so there may be things that just don't make sense or a couple parts of the code that just aren't used, stuff like that. Oh and the specular reflections in the shaders are just wrong. I'm not a big user of const correctness, something I'm working on. =) Other than that, enjoy! FYI: you will need to source the data files for the mnist dataset yourself. I wasn't sure what the license was on that, so I did not include them in the repo. You can google around and find it I'm sure. Filenames are: train-images.idx3-ubyte train-labels.idx1-ubyte Just get those files and put them in the mnistdata directory. Code can be found at: https://github.com/K9Megahertz/VisualNN