3 ms·
I don't believe "compositionality" is a serious obstacle. It is a different issue than generating an image based on a bag-of-words, so it isn't surprising that
by powera 4y ago
I don't believe "compositionality" is a serious obstacle.
It is a different issue than generating an image based on a bag-of-words, so it isn't surprising that an attempt to solve that issue didn't immediately solve the other.
But a variety of approaches can easily solve this problem.
- ummonk 4y agoYes, especially when machine translation seems to handle it just fine.
- goatlover 4y agoDoes it really, though?
- emiliobumachar 4y agoMostly. See this for five examples using Google Translate: https://www.datasecretslox.com/index.php/topic,7588.msg300075.html#msg300075 https://www.datasecretslox.com/index.php/topic,7588.msg30007...
- goatlover 4y agoI'm not sure that machine translation demonstrates compositionality, since it's translating from phrases already composed in one language to another. It only does so if understanding composition is necessary for language translation. Whereas carrying on a meaningful conversation does require understanding of how words are being put together as the conversation evolves. Thus why the Turing Test hast been considered important for determining whether an AI has achieved human-level abilities, at least as far as language use is concerned.
- ummonk 4y agoI don't see why translating from one language to relationships in art (visual language if you will) is qualitatively different from translating from one language to another.
- jameshart 4y agoRight - 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'?