Twitter | Search | |
Roy K
Data scientist. Physicist emeritus.
2,281
Tweets
288
Following
406
Followers
Tweets
Roy K Sep 19
Replying to @DrLukeOR
I'm both a competition skeptic and optimist. I agree that the winner is unlikely to be the best model. But I think these kinds of competitions have led to better understanding of which approaches are good for related tasks. Not to mention potential future transfer learning.
Reply Retweet Like
Roy K retweeted
Chris Albon Sep 17
I'm looking for three data scientists to join my team! Fun fact: A few people online have been *literally* angry at me and for this job description.
Reply Retweet Like
Roy K Sep 15
Replying to @djpardis
Linux or non-Linux?
Reply Retweet Like
Roy K Sep 11
Replying to @joelgrus
Will you sign my Jupyter notebook?
Reply Retweet Like
Roy K Sep 5
Replying to @chrisalbon
You need to try the Ethernet. Your machine can do Ethernet, right?
Reply Retweet Like
Roy K retweeted
Arundo Analytics Aug 27
Lead Data Scientist Jo-Anne Ting and Data Scientist Pushkar Kumar Jain talk about some of the problems you might face when trying to apply ML in industrial applications and how you should solve it.
Reply Retweet Like
Roy K Aug 25
Replying to @chrisalbon
I thought Wikipedia was pretty clear that no one will ever know the answer
Reply Retweet Like
Roy K Jul 29
Replying to @djpardis
What role does the DS play here? ML dev, data analysis, experimentation, etc, all of the above?
Reply Retweet Like
Roy K Jul 20
Replying to @RakeshAgrawal
Reply Retweet Like
Roy K Jul 11
Replying to @djpardis
Seems like I should have made the trip to SciPy this year...
Reply Retweet Like
Roy K Jun 28
Replying to @vboykis @gztstatistics
The next level is telling the server that your uncle will be coming over and that the server should pretend to accept his card, but should really pay with yours.
Reply Retweet Like
Roy K Jun 28
Only is allowed to call Drew short
Reply Retweet Like
Roy K Jun 24
Replying to @peteskomoroch
Here's a paper we recently posted about using synthetic data + transfer learning for anomaly detection style problems for time series
Reply Retweet Like
Roy K Jun 24
Replying to @peteskomoroch
So we have to think about the cost of synthetic and real data. Typically we start with a small real data set and go from there.
Reply Retweet Like
Roy K Jun 24
Replying to @peteskomoroch
The fundamental question we face is the trade-off between cost of data generation and performance. I.e. can we use "cheap", simple to generate data and get the performance we want?
Reply Retweet Like
Roy K Jun 24
Replying to @peteskomoroch
Yes and typically yes, assuming you created "good" synthetic data and choose the right NN architecture for the problem.
Reply Retweet Like
Roy K Jun 24
Replying to @tinymakesthings
I approve
Reply Retweet Like
Roy K Jun 24
Replying to @peteskomoroch
My team works on problems where high quality labeled data is relatively expensive to obtain, so we are focusing on synthetic data as the basis of transfer learning, then using real data for fine tuning. More constrained problems are certainly more amenable though.
Reply Retweet Like
Roy K May 27
Replying to @DynamicWebPaige
Where do you get kolache in the Bay area? (This is MTV, right?)
Reply Retweet Like
Roy K May 24
Replying to @roycoding
He told me about encountering Einstein at IAS. Despite Gell-Mann's love of language and linguistics, he declined to indulge my request for an imitation of Feynman's famous accent 😎
Reply Retweet Like