Taking derivatives is usually uncreative & tedious. Software can do it automatically.
— Reza Zadeh (@Reza_Zadeh) December 11, 2019
OTOH, in reverse, Integrals are *much* harder, need creativity, & challenge symbolic computation.
Surprisingly, a simple Transformer can learn to integrate really well. https://t.co/199avwICgO pic.twitter.com/aiVG4YyI6N
The next Kaggle reading group paper is "On NMT Search Errors and Model Errors: Cat Got Your Tongue?" by Stahlber & Byrne @emnlp2019. (@mgalle tweeted about it a while ago and it's been on my list ever since.)
— Rachael Tatman (@rctatman) December 11, 2019
Come join in tomorrow @ 9:00 AM Pacific! 😊🔖https://t.co/jYHYcXyhd5
Topic modeling allowed journalists to discover these stories:
— Rachel Thomas (@math_rachel) December 9, 2019
- insulin pumps have most reported problems in FDA database
- spinal cord stimulators help some patients, injure others
- FDA permits "export only" medical devices not for US use#NeurIPS2019 #WiML2019 @XandaSchofield pic.twitter.com/CMlDazZHUy
Using Amazon product reviews & FDA data to identify recalled & unsafe food products. @ensoesie @black_in_ai #NeurIPS2019
— Rachel Thomas (@math_rachel) December 9, 2019
Code & data process here:https://t.co/b8uoqJKxT1 pic.twitter.com/lel5odv5ih
As promised, we have made the Text-To-Text Transfer Transformer (T5) models much easier to fine-tune for new tasks, and we just released a Colab notebook where you can try it yourself on a free TPU!
— Adam Roberts (@ada_rob) December 9, 2019
👇 https://t.co/l6eI88Ih3f
(1/3) pic.twitter.com/LU5scmhbIs
Our team will be presenting Novograd at Beyond First Order Methods in ML and NeMo at MLSys workshops.
— Chip Huyen @ NeurIPS (@chipro) December 7, 2019
I'll be on a panel with some academic heroes of mine to discuss research vs industry at NewInML, 3pm Monday.
Hit me up if you want to chat! #NeurIPS2019https://t.co/WHzcfh7pG0
We now have a distilled multi-lingual model in 🤗Transformers! https://t.co/xMRRNuwaZb
— Thomas Wolf (@Thom_Wolf) December 6, 2019
This is a really cool work from UberAI on a tough question:
— Thomas Wolf (@Thom_Wolf) December 6, 2019
Is it possible to control the generations of an unconditionally trained language model?
We loved it so much that we added it to our repo and made an online demo to play with it!
Give it a try👉 https://t.co/wsoNIVGbL8 https://t.co/CNv293W0Fz pic.twitter.com/dkcurlIUXJ
Plug and Play Language Models: a Simple Approach to Controlled Text Generation
— hardmaru (@hardmaru) December 6, 2019
The demo is really cool! But they also discuss the ethics around the generation of “sensational” disinformation. An example of a “clickbait” story about freedom and democracy:https://t.co/56JRzeziCr https://t.co/PV3hCgd4ky pic.twitter.com/LNCOAcHpMI
AI Dungeon 2, An AI generated text adventure built 1.5B param GPT-2.
— Ben Hamner (@benhamner) December 6, 2019
Very creative use of text generation, and the short snippets generated by OpenAI's GPT-2 are surprisingly reasonable in this context https://t.co/sorPGVjwI8
Over the past 5 months we’ve been experimenting how to steer a mammoth without training it.
— Rosanne Liu (@savvyRL) December 5, 2019
Here is how.https://t.co/Bae0RIeSlZ
SimpleBooks is a longterm dependency dataset that is 90% the size of WikiText-103 but has 1/3 vocab and 1/4 OOV. I created it last year to test, benchmark, & do tutorials for word-level language models but didn't publish it bc small datasets get 0 love 😅 https://t.co/3TNA2xoz5Z https://t.co/8zCdf1ovdd
— Chip Huyen @ NeurIPS (@chipro) December 2, 2019
