Experimenting with Visual Identification

In pairs, you will experiment with a process used in AI to define data by features. Features are used by AI algorithms to classify data through similarities and differences with other data in a dataset.

Activity

  1. Pair up in groups of two. One person is designated Student A and the other Student B.
  2. Student A will choose an image of a single animal, and define one attribute of the image (e.g. pink nose, etc.).
  3. Student B will find an image online that matches the attribute Student A chose.
  4. Determine if they are the same type of object (not necessarily the same image, however).
  5. Student A will continue to add attributes until the objects match. Students will keep track of how many attributes were necessary in order to make the objects match.
  6. After the objects match, switch roles and repeat the experiment. As you continue the experiment, improve upon selecting attributes that are the most useful.

A simple example of the activity

You start with these two winged objects:

Your first attribute for distinction may be size. Although this is a starting point for these objects, what if, instead of an airliner, it’s a paper airplane?

Now the size distinction is irrelevant. Next, you choose the attribute movable wings.

The introduction of a butterfly causes yet another attribute to be necessary to distinguish between butterflies and birds. This process would continue until the two images are of the same category.


Feature selection is important in defining what distinguishes one object/concept from another.

Object X is this and Object Y is that because they differ in these ways (features a, b, c,...).

Class Discussion

  1. How many iterations of feature selection were necessary?
  2. Are the features as relevant if you consider any and all possible objects?
  3. Apply the strategies discussed to visually ambiguous pairs of objects. Given this Venn diagram of Monarch and Viceroy butterflies, how might you construct features that disambiguate the pair?
  4. How is this similar to the 20 Questions activity? What implications does 20 Questions have for disambiguation? HINT: Think about optimal feature selections.