Ke Tran (@ketran) header
Ke Tran (@ketran) profile picture

Computational linguist, machine learning scientist, cellist, random generator at @AmazonScience

Berlin, Germany
Joined February 2009
Photo by Ke Tran (@ketran)
You want to train BERT/RoBERTa for your language in a day or two and you have only 1 GPU. Come to DeepLo workshop, #emnlp2019 to learn more about our work on transferring entire English pretrained models to foreign languages.
5
17
1
156
Are these enormous English language models or multilingual 🤔?
CALL FOR TASKS CAPTURING LIMITATIONS OF LARGE LANGUAGE MODELS We are soliciting contributions of tasks to a *collaborative* benchmark designed to measure and extrapolate the capabilities and limitations of large language models. Submit tasks at github.com/google/BIG-Bench #BIGbench
1
1
Me: In this work, we bake a bread. Reviewer 2a: Reject! Snorting coffee via ears is not novel. Reviewer 2b: Wearing mask is not a good motivation for using ears. Also you can't snort two lines at the same time. Useless! Me: ... 🤦‍♂️
16
gained two more happy days by deciding to read the reviews on Sunday.
1
22
amsterdam is an amazing city :p
Looking for a postdoc position to work on machine translation? Topped with an exciting city and awesome colleagues? I'm hiring! uva.nl/shared-content/uva/en…
Very nice analysis of recurrent vs non-recurrent architectures. I really enjoy reading the paper arxiv.org/abs/1808.08946. Glad to see our speculation are empirically confirmed.
A very nice paper for those interested in Transformers for NLP (and if you are not, you should!) Give insights on why these models improve on high level metrics like BLEU/ppl. I like that they went the extra mile to get a real comparison to the (interesting) results of @ketran!
1
8
neat job! I'm looking forward to sentence level translation :) teapawt.org/GuillaumeLample/status…
1
2
Replying to @c_monz
@c_monz's 2006 paper stands the test of time according to Google @ILPS_Amsterdam scholar.google.com/citations…
3
torch implementation of EVE: stochastic gradient descent with feedback github.com/ketranm/optim/blo…
code for unsupervised neural HMM is available @ybisk github.com/ketranm/neuralHMM
3
how to train your HMM in neural style, unsupervised manner arxiv.org/abs/1609.09007
1
I keep discovering cool stuff from '70s and '80s. Today: Alias Method, generating samples from Multinomial distribution in O(1). – at Los Angeles, CA
doing hard attention for NMT is fun, but it's a really bad idea. – at Amsterdam, The Netherlands