Brian is a master of science student at Universidade Presbiteriana Mackenzie, Brazil. After getting his bachelor’s in pure mathematics at the same university, his love for the field has put him on a journey toward computation and neural networks. Currently he is working on a CA and a system of difference equations that could model the neuronal activity of the mammalian visual system when subjected to light stimuli. Brian is pursuing a PhD abroad, so the Wolfram Summer School will surely help him with the knowledge he needs.
Project: Optical Character Recognition with Neural Networks
This project aims at developing a useful optical character recognition by means of a neural network. It is planned that the network will be able to understand every glyph in the world—that is, when a glyph is given as input, the net should give the correct character as output. For that, a set containing all glyphs should be provided so that we can apply many image transformations, like those in images, to train the net with a very comprehensive training set.
The net is a feedforward neural network with 100,000 neurons with dot product architecture and a sigmoidal activation function, which will be fed with those glyphs throughout encoding to a tensor. So those glyphs will be transformed from images to tensors. This is the main reason why there must be many neurons—the input is probably a big tensor, say 300×400. By feeding in many fonts, it is likely that the net will recognize handwriting and other variations.
Favorite 3-Color 2D Totalistic Cellular Automaton