A short YouTube talk (recorded discussion) about practical considerations for training neural networks directly in web browsers. Speakers note constraints of browser environments (e.g., inability to open tens of thousands of HTTP connections) and propose solutions such as combining many images into a single sprite sheet and slicing it in memory to feed training (demonstrated with MNIST/handwritten-digit examples).
We'll also learn about some of the different considerations that you need to take into account when you're training in the browser. So, for example, mnist has tens of thousands of images, and if you're going to be opening HTTP connections tens of thousands of times to download those images, that could be problematic.
So, unlike training in a Python, where you can load 10,000 images one after another, you can't do that in a web browser.
Exactly. So there's a little trick that gamers actually using, game developers use, where you combine all the images into a sprite, and so you have one big sprite sheet that you download and then slice that up in memory and do the training off of that. So we'll learn actually how to do that this week.
And I think one of the coolest things you get to do at the end of this week is a project where you really train a neural network to have it recognize handwritten digits, and you really draw a character yourself, a digit from 0 to 9, and have your model run right there in your web browser to try to classify what digits you just drew.
