It’s a relatively quite day. The biggest news is probably the launch of a new Kaggle competition with a 1-GB model size restriction; a brand new approach AFAIK.
Some changes to the site:
- Add a logo image for social media sharing.
- Update permalink pattern.
- Make the site title clickable.
2nd Youtube 8M dataset competition
New @GoogleAI competition launch! This competition is the 2nd Kaggle competition based on the @YouTube 8M dataset, and is focused on learning video representation under budget constraints. https://t.co/lagtraLwA7 pic.twitter.com/FEcs1W8P9R
— Kaggle (@kaggle) May 23, 2018
Kaggle's 2nd video understanding competition challenges you to learn a model < 1 GB. Will be interesting to see the implications on kaggle's culture of ensembling: Will we see ensembles of sparse models or a single very deep model? https://t.co/zJuDJlQkh8
— Sebastian Ruder (@seb_ruder) May 23, 2018
Defining Data Science (Again)
Data scientists define their work as “gaining insights” or “extracting meaning” from data. That is way too vague.
— Miguel Hernán (@_MiguelHernan) May 6, 2018
We propose that the contributions of #datascience can be organized into 3 classes of tasks:
1. description
2. prediction
3. causal inferencehttps://t.co/8TnFOzS3Tc pic.twitter.com/iADCxZSH0f
“For causal questions, we need data and a good algorithm, but we also need expert knowledge.”
— Miguel Hernán (@_MiguelHernan) May 21, 2018
which means that reducing #causalinference from observational data to reinforcement learning algorithms (learning by trial and error) is generally impossible.https://t.co/wI0kb8g5yI
Replicating Interaction Effects
We need more papers like this. Interaction effects or more complex than main effects, and can more easily go wrong and be more dependent on model assumptions. https://t.co/Y47lczImCf
— Frank Harrell (@f2harrell) May 19, 2018
Interpretable Social Science
Interpretable social science: Find the words that really indicate something, avoiding topic correlations and confounds—Reid Pryzant, Kelly Shen, @jurafsky, Stefan Wager #NLProc #NAACL2018 https://t.co/dWHtajAhSK pic.twitter.com/x0QaPur9py
— Stanford NLP Group (@stanfordnlp) May 23, 2018
AI is Harder than You think
I agree that existing ML/AI systems focus on closed worlds. This is the fundamental reason that these systems are not safe to deploy in high-stakes open-world applications. But the idea that knowledge engineering will avoid these problems is puzzling. 1/ https://t.co/FqD5OK8IEL
— Thomas G. Dietterich (@tdietterich) May 20, 2018
On Facial Recognition By Government
(Part of a long thread. Please click on the tweet to read the full thread.)
Do not allow yourself to fall into the trap of justifying systems that enable vast human rights violations on the basis that the systems will only be used to target "bad people." Who defines "bad people," and are you sure you aren't one of them?
— 🌶 kade 🌶 (@onekade) May 23, 2018
On “Powered by AI”
(Part of a long thread. Please click on the tweet to read the full thread.)
It seems like every company is tripping over themselves in a rush to say their software is “powered by AI.”
— Jensen Harris (@jensenharris) May 23, 2018
But saying “powered by AI” is like saying you’re “powered by the internet” or “powered by computer code." By itself, it means nothing.
Here’s how I think about it:
Miscellaneous
I bought a car today, and the dealership had me check off — with a pen, on paper — that I’m not a robot. pic.twitter.com/x6nJ68e6uj
— Marci Robin (@MarciRobin) May 20, 2018
Tom the Dancing Bug: Our Nation's Leaders Analyze the Data on USA's Gun Violence https://t.co/t4KwSmQmpC pic.twitter.com/wBhSMBkoDH
— Cory Doctorow (@doctorow) May 23, 2018
honest programming books pic.twitter.com/f7ML4O5hAR
— Fred Hebert (@mononcqc) May 24, 2018
more honest books (including my own) pic.twitter.com/Eo3OTUfjyB
— Fred Hebert (@mononcqc) May 24, 2018