User:Alexander Roidl/algorithmicthesis: Difference between revisions

Revision as of 17:03, 10 October 2018

Set of experiments on generating the thesis topic by a set of rules / with a game.

Method: Print | What: Collection | Action: transform

There is a huge collection of images on http://image-net.org/ It is an attempt to classify and categories images into a huge database

I picked a random set of images:

Clouds & sky | 1399 pictures

A visible mass of water or ice particles suspended at a considerable altitude

Interestingly the set also features some cloudy explosions.

So I decided to run it through a image detection software. As I result I figured out that the software would predict quite similar values.

This is interesting as in my personal (human) opinion these images feature radical different objects.

So what would the machine see if I labeled the images myself.

It didn’t change its opinion a lot, but still the image detection wouldn’t output the same result

PREDICTED CONCEPT  PROBABILITY  nature 0.989  landscape 0.988  sky 0.987  desert 0.962  travel 0.952  outdoors 0.942  no person 0.933  summer 0.923  volcano 0.922  sun 0.917  sand 0.916  tree 0.913  soil 0.909  cloud 0.907  fair weather 0.902  hill 0.893  hot 0.891  mountain 0.884  heat 0.875  eruption 0.863 

PREDICTED CONCEPT  PROBABILITY  nature 0.993  landscape 0.991  sky 0.990  travel 0.954  desert 0.951  summer 0.948  outdoors 0.943  sand 0.940  soil 0.939  cloud 0.935  volcano 0.935  tree 0.926  sun 0.921  fair weather 0.916  mountain 0.915  hill 0.911  rock 0.910  no person 0.899  eruption 0.885  lava 0.882 

@@ Line 31: / Line 31: @@
 [[File:Screen Shot 2018-10-10 at 16.56.41.png|500px|frameless]]
-So would the machine see if I labeled the images myself.
+So what would the machine see if I labeled the images myself.
 It didn’t change its opinion a lot, but still the image detection wouldn’t output the same result