The Holes That Learning Fills
A cool visualization showing how a raw, unsmoothed n-gram only knows the word-sequences it has counted, leaving holes everywhere else; a neural model learns a smooth pattern that fills them in. Slide from memorize to learn, and drop the coverage to watch counting fall apart where learning degrades gracefully. Built in p5.js.
p5.jsinteractivemachine-learningneural-networkseducational
What’s this?
This is a pretty cool one. I made this visualization just showing the difference between learning and counting in an n-gram model.
For most dramatic effect, hit fullscreen, start with both sliders fully cranked to the end and then drop the Learn slider to see the gaps.