2. Put yourself in the shoes of a master student who has to start from scratch with your code:
— Thomas Wolf (@Thom_Wolf) January 14, 2020
- give them a ride up to the end with pre-trained models
- focus examples/code on open-access datasets (not everybody can pay for CoNLL-2003)
I often meet research scientists interested in open-sourcing their code/research and asking for advice.
— Thomas Wolf (@Thom_Wolf) January 14, 2020
Here is a thread for you.
First: why should you open-source models along with your paper? Because science is a virtuous circle of knowledge sharing not a zero-sum competition pic.twitter.com/x16jgKmLFr
I wish people stopped using the area under an ROC curve as a performance metric.
— Emmanuel Ameisen (@mlpowered) January 13, 2020
ROC curves have their use, but I find them to be most helpful accompanied by operational thresholds and other plots such as a calibration curve.
Aesthetics are important, but the main point of a figure is NOT to make the paper look pretty. When adding a figure to a paper, think hard about how the visual cues help the reader understand a result. Published network hairballs, for example, rarely covey anything useful to me.
— Rafael Irizarry (@rafalab) January 10, 2020
The problem though with "you can always add those tricks to get the numbers up" is that *very* often I see papers that don't do data aug, or don't tune hyper-params, etc, then claim their new idea helps. But then I find it's actually just a poor proxy for the things they skipped
— Jeremy Howard (@jeremyphoward) January 9, 2020
DL is methodology: building a model by assembling parameterized modules into (possibly dynamic) graphs and optimizing it with gradient-based methods.
— Yann LeCun (@ylecun) January 9, 2020
In that sense it not "sufficient" because the architecture, the objective, and the optimization method must all be instantiated.
It looks obvious now that your ML muscles are better developed when you compete with 100-1000+ teams that you have at @kaggle and not in conference type competitions where you have 3-10.
— Vladimir Iglovikov (@viglovikov) January 8, 2020
But at a time it was a big surprise to me 😀
Life is extremely fragile. It's too easy to take good health for granted.
— François Chollet (@fchollet) January 8, 2020
Am I off-base in thinking it's weird that there are so many NLP tutorials for translation? Like, approximately nobody needs to make MT models, and the people working on MT know who they are and don't need tutorials. I guess people use them for MT-analogous seq2seq tasks?
— Matthew Honnibal (@honnibal) January 8, 2020
This reminds me of great, low-tech advice from @xkcd:
— Jenny Bryan (@JennyBryan) January 7, 2020
“If I added a zero to this number, would the sentence containing it mean something different to me?” If the answer is “no,” maybe the number has no business being in the sentence in the first place.https://t.co/sEkYHq0qx8 https://t.co/owAmHcimTC
Our field isn't quite "artificial intelligence" -- it's "cognitive automation": the encoding and operationalization of human-generated abstractions / behaviors / skills. The "intelligence" label is a category error
— François Chollet (@fchollet) January 7, 2020
Techno-utopian mythologies promote social media as a fundamentally democratizing force, yet social media has enabled charismatic, personality-centered modes of authoritarianism, and micro-celebrity practices are aligned w/ anti-progressive & conspiratorial politics. -- @beccalew pic.twitter.com/wXQR0hkwxs
— Rachel Thomas (@math_rachel) January 5, 2020
