The distinction between Supervised / Unsupervised / Self-Supervised ML is illusory. What matters is *high-quality* data that captures structure in the world. Some data capture methods (video cameras, robots) can capture that structure more cheaply than others (MTurk forms).
— Eric Jang (@ericjang11) May 27, 2019
It never ceases to amaze me how much of a blind spot scientists have towards one particular subject: scientific research itself. Most of their impressions about how it works are based on EXTREMELY small samples, oftentimes on N=1: their own personal experience.
— Bojan Tunguz (@tunguz) May 26, 2019
Good advice for those pursuing a career in machine learning research https://t.co/WnqWqZTkWk pic.twitter.com/LQ9FKxjz9m
— hardmaru (@hardmaru) May 24, 2019
The issue isn't deep fakes or autogenerated text. The issue is that not a single bit of our media is digitally signed. We sign software to ensure updates to our machines haven't been tampered with unexpectedly but don't have any similar process for updating fellow humans. https://t.co/9llXcewB8j
— Smerity (@Smerity) May 23, 2019
There is an ongoing misconception that AI/ML are intrinsically valuable, and that therefore working in the field is bound to make you rich.
— François Chollet (@fchollet) May 19, 2019
A ML model is only as valuable as the problem it solves. ML without an application isn't worth anything (beyond intellectual curiosity).
100 times this
— Radek Osmulski (@radekosmulski) May 18, 2019
The flip side of this is that you should leave a job once you stop learning, which for most of the jobs is about a year.
If you can find an organization that allows you to continue to grow, that is a strong signal you might be at a great place 🙂 https://t.co/cTiukjwED3
AWD-LSTM benefits from all the work done on regularization by @Smerity . Not sure there's the same richness of regularization available just yet for transformer architectures? It's particularly important for small datasets
— Jeremy Howard (@jeremyphoward) May 17, 2019
A worrying example of the fallacy of big data.
— Neil Lawrence (@lawrennd) May 14, 2019
There are 'countably infinite things' that can possibly happen and just 500,000 cars ... around 25% of which are driving around California.
We need a lot more than just big data to have safe autonomous vehicles. https://t.co/VstFIAB0ZF
Answer to Why do many IMO Olympians fail to become mathematicians? by Niles Ritter https://t.co/SNqkS72fke
— Eric Jang (@ericjang11) May 12, 2019
I agree completely, although I keep insisting the problem is the uncoupling of data collection and analysis. I use randomized algorithms as examples of fully automating both to provide safe(r) tools. https://t.co/gBWPppDgCU
— John Myles White (@johnmyleswhite) May 12, 2019
Productivity/life tip of the day: never in my life have I regretted taking a nap.
— Tatyana Deryugina (@TDeryugina) May 11, 2019
Super useful talk from @emdodwell on the iterative process of model selection, implementation and communication #rstatsnyc pic.twitter.com/3C6mAVrKCf
— Brooke Watson (@brookLYNevery1) May 10, 2019
