The referenced link to the original work on Facebook code looks more suitable for HN:<p><a href="https://code.fb.com/ai-research/unsupervised-machine-translation-a-novel-approach-to-provide-fast-accurate-translations-for-more-languages/" rel="nofollow">https://code.fb.com/ai-research/unsupervised-machine-transla...</a>
Heads up for those of you obliging Forbes' forced resistance against adblocking:<p>There's (edit: what appears to be) an active exploit in their ad network, one that's getting around Chrome's redirect blocking through an apparent 0day.<p><a href="https://imgur.com/a/sRIB7pn" rel="nofollow">https://imgur.com/a/sRIB7pn</a><p>I'm on Chrome Beta 69.0.3497.53 on Android, so this may not apply outside that.<p>Chrome team: <a href="https://bugs.chromium.org/p/chromium/issues/detail?id=879938" rel="nofollow">https://bugs.chromium.org/p/chromium/issues/detail?id=879938</a>
> Instead of giving the system whole words, they give the system the word in parts. For example, the word “hello” might be given as 4 word parts “he” “l” “l” “o”. This means we could learn a translation for the word “he” without the system ever having seen the word “he”.<p>Can anyone add context to this? Can't seem to wrap my head around this part. Doesn't "he" as a part of a word translate differently in different words?
>> For example, “translate” between neural activity in the brain to videos on a screen,<p>That sounds almost to good to be true. Excited to see what gets developed with these techniques!
This seems like it would only work for similar languages (i.e. romance languages), in that it depends on the embeddings within languages to be similar.