Tutorial 3: Unsupervised Learning

This tutorial will quickly demonstrate how to let Recog learn on its own. Before starting, choose the Unsupervised option in the Learning menu.

  1. Create a New Character Net from the File menu. Select a width and height of 5 pixels. This will favour a square aspect ratio.
  2. Draw an arrow pointing right, then digitize and recognize it. In Unsupervised learning mode it automatically adds a new character. Since it doesn't know the name of what you drew, it will assign the character a temporary label. Draw arrows in the other three directions. Each time it will add new characters.
  3. Keep drawing, digitizing and learning these four arrows. It should start to consistently match all right arrows with the label ?01, and all left arrows with ?02, and so on. Each time you click the Recognize button, it will determine if this is something entirely new (and create a new character) or if this matches something it has seen before (an learn the pattern onto that character). Isn't this easier than supervised learning?
  4. The drawback is that it has now way of knowing what the "real" names for these symbols are. Once it has seen a few examples you have to tell it. Double click on ?01 on the list, and click the Rename button in the Info window that pops up. Rename it to Right or East. Do this for the other three arrows.

It is entirely possible that it may create two separate patterns for an arrow. This is OK; just call one East1 and the other East2. It is the boss, if it distinguishes between two isotopes of a character, it may do so. We can ignore this difference and treat both as the same character.