Find Hidden Patterns in Kandinsky Figures

For further questions and to submit your results please contact directly: heimo.mueller@medunigraz.at
For general information about the Project KANDINSKY Patters see: https://human-centered.ai/project/kandinsky-patterns

Challenge  1 – Objects and Shapes

In challenge 1 the ground truth for the pattern is known by the fact that objects are arranged on big shapes same as  the object shapes, with the restriction that in a big shape of type x, no small object of type x exists. Furthermore big square shapes only contain blue and red objects, big triangle shapes only contain yellow and red objects and big circle shape contain only yellow and blue objects:

Ground Truth of Challenge 1

Task 1: Create a machine learning algorithm that can classify Kandinsky Figures according to this ground truth.

Task 2: Identify layers and regions in the network which correspond to ”small” and ”big” shapes and the restrictions on object membership and color.

Challenge  2 – Nine Circles

In challenge 2 each Kandinsky Figure consists of 9 circles arranged in a regular grid. On Github you can find data sets for the ground truth,  for Kandinsky Figures not belonging to the pattern and Kandinsky Figures which are ”almost true”, i.e. they fulfil a hypothesis similar to ground truth, but are counter factual.

Ground Truth of Challenge 2

Task 1: Explain the Kandinsky Pattern algorithmically, i.e. train a network which classifies Kandinsky Figures according to the ground truth.
Task 2: Explain the Kandinsky Pattern in natural language.

For a specific hypothesis you can generate and test Kandinsky Figures with this Validator.  To generate images that can be correctly processed by the validator, make sure that the background color is (150, 150, 150) and the the image is 600 by 600 pixels in png format.

Challenge  3 – Blue and Yellow Circles

In challenge 3 each Kandinsky Figure consists of equal sized blue and yellow circles. On Github you can find data sets for the ground truth,  for Kandinsky Figures not belonging to the pattern and Kandinsky Figures which are ”almost true”, i.e. they fulfil a hypothesis similar to ground truth, but are counter factual.

Ground Truth of Challenge 3

Task 1: Explain the Kandinsky Pattern algorithmically, i.e. train a network which classifies Kandinsky Figures according to the ground truth.
Task 2: Explain the Kandinsky Pattern in natural language.

For a specific hypothesis you can generate and test Kandinsky Figures with this Validator.  To generate images that can be correctly processed by the validator, make sure that the background color is (150, 150, 150) and the the image is 600 by 600 pixels in png format. The circles are 30 pixels in diameter

Additional information can be found at https://arxiv.org/abs/1906.00657   All datasets can be downloaded at https://github.com/human-centered-ai-lab/dat-kandinsky-patterns