Digit Draw: Neural Network Digit Recognition
MrEconomical (2191)

Neural Network Recognition

Digit Draw is a program to recognize 28x28 pixel hand-drawn digits. It uses a neural network made with brain.js trained on MNIST data (a database of hand-drawn digits). It takes in an array of 28x28 for the darkness of each pixel and outputs an array of length 10 of the certainty for each digit. The neural network works fairly well and is able to predict the digit most of the time.


Input Size: 784
Hidden Layers: 1
Hidden Layer Neurons: 392
Activation Function: Sigmoid
Output Size: 10


brain.js - https://github.com/BrainJS
MNIST - http://yann.lecun.com/exdb/mnist/

You are viewing a single comment. View All
AdCharity (1307)

nice project :D although sadly I believe this has been done on repl before (with very similar specifications - not saying you're copying its very common to use MNIST data as a somewhat first project)

Edit: just a question, but can brain.js train on base 64 images? I was going to try but I got lazy

MrEconomical (2191)

@AdCharity no the input must be array of number or object for a feed forward back propagation network I tried with base64

AdCharity (1307)

@MrEconomical ah >:l
btw that is a mad face. Btw if I normalized base64 images (like if one was longer than the other and cut of the ends) then would it work? Brain.js can accept strings extremely quickly if you encode the data.... mmm