We’ve reached double-digit number! Tons of good tweets for you today. Read on! :metal:
(For those wondering the meaning of the title, just ctrl+f “Bayesian”. :grin:)
Nature Machine Intelligence Boycott
New Guardian op-ed on the Nature Machine Intelligence boycott. https://t.co/iufWybSozQ
— Neil Lawrence (@lawrennd) May 29, 2018
Why Thousands of Researchers Are Boycotting Nature’s Upcoming AI Journal https://t.co/T70losN2JO via @gizmodo
— Bojan Tunguz (@tunguz) May 30, 2018
Landmark Retrieval (Kaggle)
1st place solution:
1st Place in the @kaggle Google Landmark Retrieval Challenge!! All the hard work finally paid off 😊
— Mikel Bober-Irizar (@mikb0b) May 30, 2018
Read in-depth about our solution here: https://t.co/AhcQXX4GQl pic.twitter.com/BMy3LurKr8
Semantic Search on Code
(How To Create Natural Language Semantic Search For Arbitrary Objects With Deep Learning)
I am excited to share research my colleague Ho-Hsiang and I have done that uses deep learning to build semantic search on code. We provide a fully reproducible example. Amazing demo of #MLonCode @Github. Thanks to @fastdotai for the tools/ideas! https://t.co/a76ejIOwVC pic.twitter.com/9euXbTxeJq
— Hamel Husain (@HamelHusain) May 29, 2018
Play Games by Watching Youtube
Playing hard exploration games by watching YouTube (DeepMind). One-shot imitation allows an agent to exceed human-level performance on the infamously hard exploration game Montezuma's Revenge, even if the agent isn't presented with any environment rewards. https://t.co/kpsFtfJH2W pic.twitter.com/WUXzjNP4as
— hardmaru (@hardmaru) May 30, 2018
"Observe and Look Further: Achieving Consistent Performance on Atari," Pohlen et al.: https://t.co/TvIaa8HDZM
— Miles Brundage (@Miles_Brundage) May 30, 2018
"Playing hard exploration games by watching YouTube," Aytar and Pfaff et al.: https://t.co/SOJU3GuBfo
Both from (mostly) DeepMind and both claiming v. strong MR results
Diagnosing Schizophrenia
Can computers detect mental illness from human language use? Automatic Detection of Incoherent Speech for Diagnosing Schizophrenia by @dan_iter, Jong H Yoon & @jurafsky. CL & Clinical Psych #NAACL2018 workshop https://t.co/Qpt2CZIt71 pic.twitter.com/MxtdQnylG6
— Stanford NLP Group (@stanfordnlp) May 29, 2018
I am deeply uncomfortable with this study. One of the metrics used is "ambiguous pronoun usage", and using "they" to refer to a named human referent is specifically marked as an error: "Resolving they to refer to Joe would mean the incorrect pronoun is used." https://t.co/J87qqL8NxJ
— Rachael Tatman (@rctatman) May 29, 2018
Another big warning flag for me: expletive "they" would also be marked as ambiguous and, again, is a perfectly licit in dialects like African American English https://t.co/SlgmGva3Hz
— Rachael Tatman (@rctatman) May 30, 2018
Machine Learning Pratica (Keras)
Introducing "Machine Learning Practica", from Google AI: interactive courses to master the foundations of machine learning, with exercises in Colab. Already covers convnets, data augmentation, and fine tuning -- all with Keras: https://t.co/ESDRlw7YV1
— François Chollet (@fchollet) May 29, 2018
Direct link to the image classification practicum: https://t.co/0UwL19gSu9
— François Chollet (@fchollet) May 29, 2018
Attractor Landscapes
Why do systems – cultures, environments, economies – seem to get stuck, then change all at once? Here's a mini-interactive about one of my favorite ideas, attractor landscapes!
— Nicky Case (@ncasenmare) May 29, 2018
🗻 https://t.co/sDtGhxVVqB 🗻 (playing time: 5 min) pic.twitter.com/EU0OLrLGtI
Choosing Colors
I just published the next part of our "What to consider..." series: Colors. Because every single chart needs them. https://t.co/8jAesQgp49
— Lisa Charlotte Rost (@lisacrost) May 29, 2018
Applying to DS Jobs
New blog post 🎉! A collection of advice and resources for folks applying to data science jobs, starting from preparing your application to negotiating an offer. https://t.co/NADMjQeOVK
— Emily Robinson (@robinson_es) May 29, 2018
Notables
Big deal for heterogeneous data in scikit-learn (columnar data, pandas DataFrame, CSV files): ColumnTransformer just merged in development version:https://t.co/jJe5T2XWt1
— Gael Varoquaux (@GaelVaroquaux) May 29, 2018
Thanks to @jorisvdbossche, @amuellerml, Joel Nothman!
New Blogpost: Beyond Numpy arrays in Python
— Matthew Rocklin (@mrocklin) May 29, 2018
Preparing the ecosystem for sparse, distributed, and GPU Numpy-style arrays.https://t.co/RbDLFt4Fkx
NLP Architect by @IntelAI Lab
— ML Review (@ml_review) May 30, 2018
Python library for exploring SoTA Deep Learning topologies & techniques for #NLProc https://t.co/duO5yHcsDc pic.twitter.com/APXhoYHalt
Super cool @PyTorch reimplementation (+ new stuff) of our @DeepMindAI differentiable stacks/queues/etc (NIPS'15) by @Yale undergrad(!) Will Merrill. Check it out!https://t.co/1jlHp6rnL3
— Edward Grefenstette (@egrefen) May 29, 2018
Reproducibility in machine learning - why it matters and how to achieve it: great overview from @DeterminedAI https://t.co/o6NI83pnbf pic.twitter.com/bDMtbOsM8T
— Ben Lorica 罗瑞卡 (@bigdata) May 29, 2018
We ran a survey to better understand mortality in Puerto Rico after hurricane María. The official death count of 64 is likely a substantial underestimate. Lack of access to medical care was a major problem. https://t.co/PtLHVgqDTp
— Rafael Irizarry (@rafalab) May 29, 2018
Code and data are here: https://t.co/SyzXecmso1 pic.twitter.com/YX5OCt1g6i
Miscellaneous
A question that comes up after almost every Enterprise NLP talk is “How do you handle other languages?”. There are many possible approaches, but the unpleasant reality is most businesses will struggle to get even basic models shipped and working in English.
— Peter Skomoroch (@peteskomoroch) May 29, 2018
In Part III of my interview with Erik Walenza for his @IotoneHQ podcast we get into the challenges of building AI products for industrial customers, and and the importance of immediate value and trust creation. https://t.co/UbH5aJ9qkf
— Drew Conway (@drewconway) May 29, 2018
Important advice from @petewarden: "It may seem obvious, but your very first step should be to randomly browse through the training data you’re starting with. Copy some of the files onto your local machine, and spend a few hours previewing them." https://t.co/ds2oxTqpEB
— Peter Skomoroch (@peteskomoroch) May 29, 2018
Today, as a developer who has been programming for a quarter-century, I am now 20+ minutes into trying to figure out how a five-line for loop is looping infinitely.
— Pete Holiday (@toomuchpete) May 28, 2018
Just in case you needed a reminder that these things happen to everyone.
This is the kind of thing people mean when they bring up the human aspect of algorithms and automated systems.
— Mikhail Popov (@bearloga) May 29, 2018
Someone made that decision, either carelessly or intentionally, to exclude PR from US news in some config file somewhere.
The consequences of this are…profound, imho. https://t.co/XdYOEH3p1C
The @ICRC is using @Twitter data to anticipate and respond to humanitarian crises. Here's how. https://t.co/LkyWBO4NkO pic.twitter.com/3v7YoQgwIc
— Devex (@devex) May 29, 2018
Another existential risk to auditors: algorithms that neither companies nor auditors truly understand but which are vital to the business model. Auditors need to audit algorithms (including their own). https://t.co/imh3frwyA9
— Cathy O'Neil (@mathbabedotorg) May 29, 2018
False positive and false negative results for facial recognition “are likely to arise disproportionately in relation to people who belong to ethnic minorities in Australia”.
— Belinda Barnet (@manjusrii) May 29, 2018
Weapon of math destruction right there @mathbabedotorg
Facebook is inviting research proposals on crisis informatics and the role of social media in disaster response and recovery: https://t.co/taQtlDp7Py Please consider submitting a proposal!
— Lada Adamic (@ladamic) May 29, 2018
If AI is a new industrial revolution shouldn't we be... incredibly concerned? The industrial revolution was a time of great chaos and misery for millions of people, and it occurred during a period with less extreme weather and a less connected world.
— Jack Clark (@jackclarkSF) May 29, 2018
Don’t worry if your results aren’t significant.
— Nihilist Data Scientist (@nihilist_ds) May 29, 2018
It’s likely that the phenomenon you’ve dedicated your life towards studying is so noisy and hard to measure that nothing of subtance can be said about it at all.
It’s not your fault, it’s just hopeless, that’s all!
TFW a PhD student is presenting about Bayesian statistics, and they keep saying it needs a "prayer", and you don't think that's a very solid foundation for inference, and then you realise they are actually saying "prior".
— Jeffrey Rosenthal (@ProbabilityProf) May 28, 2018
No GANs were harmed in the making of this buzzword (which I may not have invented.) Regardless: Big Empiricism seems like an interesting force in AI research, though "big" may be more due to $$$ and rubbish traits of current algos rather than specific scientific innovation. https://t.co/bPBRxH8yCK
— Jack Clark (@jackclarkSF) May 29, 2018
“Director of AI”
I’m talking to so many enterprises where “Director of AI” or similar titles are held by people with absolutely no background in AI, except with an occasional recent familiarity. What’s up with that?
— Delip Rao (@deliprao) May 22, 2018
All conversations with such folks is depressing. From the conversations it appears they seem adept in name dropping and jargon flinging, but cannot clearly articulate how AI affects their business.
— Delip Rao (@deliprao) May 22, 2018
Many of these folks have actual ML phds and data scientists working for them. I wonder how depressing it must be to walk into an office and work for these namesake leaders.
— Delip Rao (@deliprao) May 22, 2018