Twitter | Search | |
Loreto Parisi
Computer Engineer. Technical Director of Machine Learning Father of an amazing girl and a cute boy.
11,743
Tweets
655
Following
889
Followers
Tweets
Loreto Parisi 4h
Replying to @quocleix
It would be worth at some point to have a timeline of credits more than citations for (for which relations graph models exists). So that anyone could get credited for a specific new outcome.
Reply Retweet Like
Loreto Parisi retweeted
Quoc Le 10h
Replying to @quocleix
First, the idea of using pre-trained language models for transfer learning was proposed in our paper in 2015: . (2/3)
Reply Retweet Like
Loreto Parisi retweeted
Quoc Le 10h
Replying to @quocleix
Second, the idea of embedding words through a language model was proposed in our followup paper in 2016: . (3/3)
Reply Retweet Like
Loreto Parisi retweeted
Quoc Le 10h
Peters et al 2017 and 2018 (ELMo) get the credit for the discoveries of 1) Using language models for transfer learning and 2) Embedding words through a language model ("contextualized word vectors"). ELMo is great, but these methods were proposed earlier. (1/3)
Reply Retweet Like
Loreto Parisi retweeted
Andreas Mueller 19h
A lightGBM-style gradient boosting in scikit-learn is ready for reviews/test driving: Any feedback welcome! Amazing work by If you use GradientBoostingClassifier/Regressor, definitely try it (if you're not a afraid of using a github branch ;)
Reply Retweet Like
Loreto Parisi retweeted
Andreas Mueller 8h
I'll take it as a sign of enthusiasm ;) also is in the lucky (IMHO) position to get paid for this. Though that means I can tell him what to do, not you ;) [Also he'd probably argue with me if I told him *not* to add categorical support]
Reply Retweet Like
Loreto Parisi retweeted
Federico Vaggi 13h
It's an open source project run by a team of amazing scientists and volunteers, largely unpaid. Telling them what they "must" do is pretty poor form.
Reply Retweet Like
Loreto Parisi 4h
errata corrige you *should* consider to add categorical features. Sorry guys it was just a suggestion. Thank you very much for your work. It’s *priceless*. Without projects like this a lot of people would actually have no opportunities in
Reply Retweet Like
Loreto Parisi retweeted
Sanyam Bhutani Feb 20
This is definitely interesting: "Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks"
Reply Retweet Like
Loreto Parisi retweeted
JavaScript Open Source Trends 🔥 www.jsdotgit.com 14h
🔥 : "NSFW detection on the client-side via Tensorflow JS" is trending today (70 ⭐️ so far)
Reply Retweet Like
Loreto Parisi 14h
You must absolutely add categorical feature to this Gradient boosting!
Reply Retweet Like
Loreto Parisi 14h
Replying to @gneubig @jeremyphoward
This means that SDE performs better than BPE (like in WordPiece/BERT)? That’s pretty cool for mixed languages models. Do you have a SDE implementation? In SentencePiece is used FastBPE in C++
Reply Retweet Like
Loreto Parisi retweeted
Graham Neubig Feb 20
Very nice summary of the problems of using subword segmentation on multilingual corpora! One motivation behind our "soft decoupled encoding" paper () was over-segmentation of less frequent scripts in our 1000-language MT system. Important stuff to know.
Reply Retweet Like
Loreto Parisi retweeted
Eclipse DL4J 15h
Contextual Word Representations: A Contextual Introduction
Reply Retweet Like
Loreto Parisi retweeted
Andrew Trask Feb 19
Papers are published online - perhaps it's time we started taking conferences online as well! More attendees (could be $30 instead of $2000+, no visa issues) More content (anyone could livestream a preso) More often (could have every quarter!) Seems like a no brainer
Reply Retweet Like
Loreto Parisi retweeted
Judit Acs Feb 19
One size fits all? Exploring BERT's 104 language WordPiece vocabulary
Reply Retweet Like
Loreto Parisi retweeted
Thomas Wolf Feb 20
Replying to @Thom_Wolf
[2/3]❓NeuralGen: A workshop on advances/evaluation of Language Generation, co-located with NAACL 2019 ⏱️ Deadline: March 11th (one week after ACL) 📣 Call for Paper: 👉 Submitting work published/submitted elsewhere is permitted (may not be reviewed)
Reply Retweet Like
Loreto Parisi retweeted
Shanqing Cai Feb 19
You can also consider training the model directly in tensorflow.js using tfjs-node-gpu, eliminating the need to port the model. When running in browser, allow user to flag false negatives so the model can be fine-tuned at client side. Just some ideas 😁
Reply Retweet Like
Loreto Parisi Feb 19
Thanks to we now have a simpler interface to the language model and some developer put the model here It’s pretty evident that the model shows nonsense sentences and biases in mostly syntactically correct expressions
Reply Retweet Like
Loreto Parisi retweeted
DataScienceNigeria Feb 17
Now you can do without code with Ludwig from Just provide a CSV file, a list of columns as input & a list of columns as output& it does the rest. It trains,predicts & visualizes. Ludwig main innovation is the idea of datatype specific encoders & decoders
Reply Retweet Like