Bayesian Learning Web Apps!

I’ve developed my first web-hosted app that features a Bayesian neural network using 2-dimensional convolution (think computer-vision) to classifies handwritten characters including the digits 0-9 and the upper case alphabet A-Z, and some of the lower-case alphabet. The neural network was trained using the extended MNIST dataset (EMNIST) which includes handwritten characters as well as the handwritten digits that compose the original MNIST dataset.

There is a sidebar that can be opened with the little arrow in the top left of the web window. There you can select how many re-samplings you want the model to perform to estimate probabilities for your handwritten digit. Once you draw a number, press submit and the model will output it’s predictions. After the model finishes predicting it outputs its probabilities for each possible digit in a bar graph below the canvas. There you can see if the model is very (or not) certain about what number you wrote. In the sidebar is a checkbox to keep track of how many correct (and incorrect) predictions the model gets for you. The next update is aimed to adding OpenCV code to perform multi-digit separation of people’s input to be able to recognize multi-digit numbers. Hope you enjoy!