Wracając do sedna dzisiejszego zamieszania z artykułem autosportu / motorsportu. Kubica mówi w nim, że podejmie swoją decyzję w ciągu kilku dni. Rozważa opcję z Williamsem lub posadę kierowcy symulatora w Ferrari. Więc za kilka dni dowiemy się co go czeka za rok. #F1PL
You May Also Like
Ironies of Luck https://t.co/5BPWGbAxFi— Morgan Housel (@morganhousel) March 14, 2018
"Luck is the flip side of risk. They are mirrored cousins, driven by the same thing: You are one person in a 7 billion player game, and the accidental impact of other people\u2019s actions can be more consequential than your own."
I’ve always felt that the luckiest people I know had a talent for recognizing circumstances, not of their own making, that were conducive to a favorable outcome and their ability to quickly take advantage of them.
In other words, dumb luck was just that, it required no awareness on the person’s part, whereas “smart” luck involved awareness followed by action before the circumstances changed.
So, was I “lucky” to be born when I was—nothing I had any control over—and that I came of age just as huge databases and computers were advancing to the point where I could use those tools to write “What Works on Wall Street?” Absolutely.
Was I lucky to start my stock market investments near the peak of interest rates which allowed me to spend the majority of my adult life in a falling rate environment? Yup.
The paper is a good example of lots of elements of good experimental design. They validate their metric by showing lots of variants give consistent results. They tune hyperparamters separately for each condition, check that optimum isn't at the endpoints, and measure sensitivity.
They have separate experiments where the hold fixed # iterations and # epochs, which (as they explain) measure very different things. They avoid confounds, such as batch norm's artificial dependence between batch size and regularization strength.
When the experiments are done carefully enough, the results are remarkably consistent between different datasets and architectures. Qualitatively, MNIST behaves just like ImageNet.
Importantly, they don't find any evidence for a "sharp/flat optima" effect whereby better optimization leads to worse final results. They have a good discussion of experimental artifacts/confounds in past papers where such effects were reported.
Fun going down this list and thinking: "Hmm, plausible at a well-run modern software shop", "Hmm, possible, but requires implausible tradeoffs", "Literally disallowed by languages", and "If you were to attempt doing that our test suite wouldn't let you merge."
I think we as an industry celebrate (not quite the right word) failure too much and don't celebrate success nearly enough. There is no DailyWTF for competent execution, word of which generally stays pretty local to the source while incompetence passes into legend.
Alrighty let me try to thread the needle on being the change I want to see in the world while not giving away anything that will get me in trouble:
Ruby has wonderful developer ergonomics. Typed languages are easier for machines to guarantee the correctness of. We built a type checker for Ruby (and I believe it is slated for OSS release sometime).
(1) The notion that R is well-suited to "building web applications" seems totally out of left field. I don't feel like most R loyalists think this is a good idea, but it's worth calling out that no normal company will be glad you wrote your entire web app in R.
(2) It is true that Python had some issues historically with the 2-to-3 transition, but it's not such a big deal these days. On the flip side, I have found interesting R code that doesn't run in modern R interpreters because of changes in core operations (e.g. assignment syntax).
(3) "Most of the time we only need a latest, working interpreter with the latest packages to run the code" -- this is where things get real and reveal some things that hurt data scientists. If this sentence is true, it's likely because you don't share code with coworkers.
(3) Really is a broader issue in data science: people only think of what they need to do their work if no one else existed and code was never maintained. Junior data scientists almost always operate on projects they start from scratch and don't have to maintain for long.