So much this. Or maybe you're indirectly tweaking regularization.
— Jeremy Howard (@jeremyphoward) August 28, 2018
This is why researchers that run experiments need to be good practitioners. If you're not training with appropriate LR and regularization, your experiments are probably meaningless. https://t.co/mTJHZOyBgC
Thank you @jackclarkSF for pointing out this massive fail.https://t.co/Mb849GHWpP
— Jeremy Howard (@jeremyphoward) August 27, 2018
Researchers: no longer can you ignore ethics issues and hope no-one will notice. Engage with them, as if they were as important as your choice of loss function. pic.twitter.com/8RdtxiHmu5
I literally had to argue with a reviewer about showing a full time series of data in a paper. They argued reporting summary statistics was enough. Nope. It really isn’t. Datasaurus rex is my favorite example for illustrating why... pic.twitter.com/eHl2KX5hV8
— Dan Weaver (@DanWeaver_ca) August 27, 2018
If they want to improve the quality of scientific publications, rather than banning p-values or changing the 0.05 threshold, journals should make us show the data. pic.twitter.com/UFz3xrPQZJ
— Rafael Irizarry (@rafalab) August 27, 2018
I like Notebooks and they can be useful even in production-grade environments (see “Beyond Interactive: Notebook Innovation at Netflix” https://t.co/BJjjPQwbpT). However, all of @joelgrus points, especially the one about hidden state, are 100% valid and should be addressed https://t.co/ANUh7wixkX
— Xavier 🎗🤖🏃 (@xamat) August 26, 2018
Some information, such as TensorFlow Prints, gets invisibly printed in the backend rather than in the notebook
— Ian Goodfellow (@goodfellow_ian) August 25, 2018
Notebooks generally don’t play very friendly with source control tools such as git
— Ian Goodfellow (@goodfellow_ian) August 25, 2018
I don’t like notebooks either. I agree with a lot of @joelgrus ‘s reasons and have a few others of my own https://t.co/D0DxniDyxt
— Ian Goodfellow (@goodfellow_ian) August 25, 2018
Yes! And this is basically the idea behind graph neural networks / relational networks (eg, https://t.co/SeihrvMIPs or https://t.co/SThMV0E8So). The whole “loopy message passing” with neural nets thing is a great idea (imo) :)
— Will Hamilton (@williamleif) August 25, 2018
If you squint, transformer is like a densely connected factor graph. Network depth approximates the number of rounds of loopy belief propagation.
— Victor Zhong (@hllo_wrld) August 25, 2018
Trust, good faith, and a cooperation mindset are the foundational layer of all relationships. Without them, you cannot have a society, or a company, etc. We should place these first, far above self-interest, since the loss of these spells the loss of everything else.
— François Chollet (@fchollet) August 23, 2018
Very well put! But I would go even further: "To be a science, data-science should start where science does." I tried to add it here https://t.co/rXgGXv58LL but one reviewer got offended. #Bookofwhy #ecobookclub #Epibookclub https://t.co/FQos7QWL4Z
— Judea Pearl (@yudapearl) August 22, 2018
