Hacker News Reader

Comment by avalys | Hacker News Reader

This is hilarious. But I don't understand how the network learned the words and color associations for "turd", "poop", "stanky" and "butt", since presumably these were not words Sherwin-Williams used for their own colors and thus would not be present in the training set all. What am I missing?

View on HN

Probably bias on the author's side where he would select those words over the nonsense outputs or the unfunny ones.

The names are done on a letter basis rather than a word basis judging from some of the outputs.

Even still, the algorithm has no knowledge of these words. The likelihood of producing them is infinitesimal. Why would it produce real words more often than gibberish, or words in every other language on earth? I don't think this argument holds.

If the algorithm is trained on English words, it will learn English phonotactics (i.e., syllable structure, what sounds are allowed at the beginning of words, etc). That makes it much more likely to generate valid words, as the set of strings that follow English phonotactic rules << the set of gibberish strings.

Agreed, but I can't imagine with a sample space of just 7700 paint names that it would generate turdly any more frequently than bvlogd or a gazillion other marginally pronounceable six letter strings, without some additional bias not mentioned in the article.

If I were the author I would train another NN that compares the color with a huge set of photo+definition and decide from there. For instance "Coral Gray" looks quite accurate when compared with the average recreational diver photo of coral reefs (but not the real coral specimens, which tend to be very colorful when correctly photographed).

Reply on news.ycombinator.com