Attempting to automate the solving of Bongard Problems (BPs).
Bongard M.M.- "Pattern Recognition" 1967
Thanks to Foundalis, my initial dataset is pulled form their representations of the original problems presented by Bongard
The index of BP images that serve as input to the solver reside in ./resources/images
The related dictionary of solutions is in .csv format in ./resources/data
- Given n images, produce a description of difference between the left and right sides of a BP
-
Generalize a set of features from available solutions for original problem-set (may be impossible - work backwards instead, limits, rules)
-
Prove generalization holds for community-generated BP's
-
Extract a large enough set of #salient_features to encompass solution #feature *** Shortcut: ChatGPT image recognition with a tuned prompt **Craft prompt for chatGPT to recognize salient features *** Test prompt for 100% accuracy on original problem set with no prior knowledge *** Test prompt for 99% accuracy on community-generated problems
-
Prove that #feature is shared among 6 left images
-
Prove that #feature is not shared in any of 6 right images