Handwritten digit recognition neural network

Description

Run on TurboWarp for it to decompress and work faster https://turbowarp.org/543530687 This project isn't intended to be anything impressive. This is the first time I managed to implement a working neural network. Neural network has dimensions: 784 x 196 x 196 x 64 x 10 Trained in program I wrote in Javascript and GLSL (WebGL 2). For training code I used this tutorial: https://stevenmiller888.github.io/mind-how-to-build-a-neural-network/ BUT some divisions had to be replaced by multiplications for it to work. Trained on EMNIST-digits dataset for 2 epoches: https://github.com/aurelienduarte/emnist/ https://arxiv.org/abs/1702.05373 One of the challanging things was fitting this neural network, which consists of 214472 32bit floating point numbers, into 5 Mb project.json limit. To achive it I reencoded binary representation of those numbers into base-64. When you start the project for the first time, it takes every base-64 character and splits it into 6 bits. As soon as there are 32 bits, it calculates mantissa, exponent and sign and combines those things together to get 32bit float. First layer alone consists of 820588 base-64 characters, so that is the reason why it takes so long for data to decompress. Press H to delete decompressed data. This is not a CNN(convolutional neural network) Keywords for the search/Please ignore: image recognition number recognition mnist emnist