3 ms·
Should be noted this idea is not that novel though - it's just replacing word vectors with a pre-trained model. Interesting that it works so well but not very s
by andreyk 8y ago
Should be noted this idea is not that novel though - it's just replacing word vectors with a pre-trained model. Interesting that it works so well but not very surprising.
- laichzeit0 8y agoIt may not be novel, but then why do you we still have commercial APIs (Microsoft, Google, IBM Watson, etc.) where there is pretty much no way to "fine tune" them to your domain with a small set of supervised examples? We all know domain adaptation is a real problem. Instead you either have to roll-your-own models in-house (which defeats the whole point of using a ready made cloud solution) or deal with whatever accuracy you happen to get from those APIs. IMHO this is an area where you can make some serious competitive headway in commoditised AI/ML. Do all the heavy lifting of pretraining and give your customers an API to "fine-tune" with. Who is currently doing this?
- pwaai 8y agoWell i hope that http://js.fo http://js.fo will be that as i add more aiml libraries although currently focused on web area. I hope to expend in other areas
- radarsat1 8y agoWait what is the difference between word vectors and a pre-trained model? Aren't word vectors generated by a neural network trained to predict the next word, or to recognize noise (noise-contrastive training)? What is the difference in a "pre-trained model" as opposed to the training needed to generate word vectors?