NLP-based workflow tool for journalists & storytellers. Style transfer for text: Augmenting writers with an AI-writing-assistant.
Bengaluru, India
Joined March 2015
- Tweets410
- Following24
- Followers29
- Likes162
DOinterview retweeted
We show how to read a human-interpretable algorithm from a neural network's weights, by studying the circuits formed by the connections between individual neurons: distill.pub/2020/circuits/zo…
DOinterview retweeted
T-NLG
"Better natural language generation can be transformational for a variety of applications, such as assisting authors with composing their content, saving one time by summarizing a long piece of text with digital assistants" @MSFTResearch #AI #NLP
microsoft.com/en-us/research…
DOinterview retweeted
The longer it takes to become successful, the more prepared you will be to handle it.
DOinterview retweeted
Totally crazy that you can squeeze BERT down to 7 MB without much drop in performance - "Extreme Language Model Compression with Optimal Subwords and Shared Projections," Zhao et al.: arxiv.org/abs/1909.11687
T-NLG
"Better natural language generation can be transformational for a variety of applications, such as assisting authors with composing their content, saving one time by summarizing a long piece of text with digital assistants" @MSFTResearch #AI #NLP
microsoft.com/en-us/research…
DOinterview retweeted
NLP leaderboards are turning into who can pretrain the biggest architecture for the longest. But hey, maybe that's a good thing - people will realize they can't compete and get back to doing good science, caring about ablations instead of absolute leaderboard positions.
DOinterview retweeted
Early on the neural network landscape was fragmented. In 2015, Keras started attracting attention. Today it’s Keras/TensorFlow and PyTorch.
(As of TF 2.0, Keras is the high level API for TensorFlow.)
3/3
DOinterview retweeted
For example:
- don't try to solve all problems with automation. humans need help from humans. use tools. but add support staff, don't cut it.
- divine the solution. customers don't know how to solve their problems. you do. tell them how -- and implement it.
An example election summary produced by the BBC system.
This story was created using some automation.
bbc.com/news/politics/consti…
DOinterview retweeted
Three weeks ago, we released a report with @PolisLSE highlighting how artificial intelligence can transform journalism.
We're building on that work with new guidelines to help newsrooms approach using #JournalismAI for the first time. goo.gle/2YENKZO
DOinterview retweeted
Read our original post about how artificial intelligence is changing the news industry: goo.gle/2KvTCPi
DOinterview retweeted
Excited we’re hosting Computation + Journalism on March 20!
Keynotes: @amandacox @dfreelon @DavMicRot
Talks: @databyler @micahcohen @AllisonSchrager @AlbertoCairo @kanarinka @ChristiansenJen @spetulla
Still accepting research, practice or tech papers! cj2020.northeastern.edu
"AI can often augment the work that journalists do. It can make people more efficient. It can enhance the quality of journalism processes.
We should think about the hybridization of journalism and automation." ~@ndiakopoulos
"Only 15% of a reporter’s job and 9% of an editor’s job could be automated using technology." ~@ndiakopoulos
DOinterview retweeted
We have open-sourced wav2letter@anywhere, an inference framework for online speech recognition that delivers state-of-the-art performance. ai.facebook.com/blog/online-…
DOinterview retweeted
Hang it there long enough
Truly listen to your customers
Build awesome software
You will catch a break – at San Mateo, CA
1/7
All The Ways You Can Compress BERT mitchgordon.me/machine/learn…
Model compression reduces redundancy in a trained neural network. Improved memory and inference speed efficiency can also save costs at scale. ~Mitchell #ML #AI #NLP
6/
Quantization - Truncates floating point numbers to only use a few bits (which causes round-off error). The quantization values can also be learned either during or after training.