3 ms·
>> This class of low-cost depth sensors emerged just as Deep Learning boosted Artificial Intelligence by accelerating performance of hyper-parametric function a
by YeGoblynQueenne 8y ago
>> This class of low-cost depth sensors emerged just as Deep Learning boosted Artificial Intelligence by accelerating performance of hyper-parametric function approximators leading to surprising advances in image classification, speech recognition, and language translation.
So I haven't really looked at machine translation benchmarks for a while, but the paper linked by the article itself (as an example of the "surprising advances" in language translation) reports results below the state of the art (from 2014) (see [1] below).
The system noted as the state of the art, [2], is a phrase-based statistical machine translation system, which basically boils down to n-grams, with a few tricks up their sleeves such as a 5-gram operation sequence model, hierarchical lexicalised reordering, an unsupervised transliteration model (for translations between Hindi and Urdu) etc etc.
This is interesting to me because I hear this claim, that deep learning has advanced machine translation significantly, very often- but when I look for results, I can't really find anything that supports that. It's obvious that machine translation research currently involves a lot of deep learning, but whether that is actually working, or it's just what people are trying at the moment to see if it works, I can't tell.
It certainly doesn't look -to me at least- like we've seen anything like the big leap on ImageNET and CIFAR results, when it comes to machine translation. The paper linked by the article is a good example: their baseline is a 33.30 BLEU score and their system raises that to 34.81, an improvement of 1.51, a less than 5% increase (on the previous result). And that's only on English-to-French (and still behind the top results).
I really don't see the "surprising advancements" here.
______________
[1] https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf https://papers.nips.cc/paper/5346-sequence-to-sequence-learn...
[2] https://www.aclweb.org/anthology/W14-3309 https://www.aclweb.org/anthology/W14-3309