3 ms·
Right - your training data set is images plus descriptions. But the descriptions are not typically descriptions of composition. Descriptions of Napoleon Crossi
by jameshart 4y ago
Right - your training data set is images plus descriptions. But the descriptions are not typically descriptions of composition.
Descriptions of Napoleon Crossing the Alps are unlikely to read 'A small frenchman wearing a silly hat riding on a horse'. So why would an AI trained on such image descriptions develop any sense for 'compositionality'?