abhgh

17 ms·

1. 151.▲
   by abhgh 3y ago

   Your comment reminded me this talk [1] by David Duvenaud where he discusses his work on Neural ODEs, introduced in the paper [2] titled "Neural Ordinary Differential Equations" (this was one of the Best Papers in Neurips 2018). Ap
2. 152.▲
   by abhgh 3y ago

   I'm genuinely trying to understand (based on this and another comment above): wouldn't storing data for pre-training vs fine-tuning carry the same risks?
3. 153.▲
   by abhgh 4y ago

   I think we are seeing Knowledge Distillation at a large scale. OpenAI trains this megamodel which has an impressive amount of real-world knowledge packed into it, and annotation tasks such as these are effectively extracting a small portion
4. 154.▲
   by abhgh 4y ago

   Seconding the recommendation for Nautilus. I used to subscribe to their print edition for a while, and had loved it.
5. 155.▲
   by abhgh 4y ago

   I wasn't aware of this! It's funny because when we were taught Conceptual Dependency (in a GOFAI course as part of my Masters), this is something we used to joke about, that so many of the examples are violent. Yes, a great loss.
6. 156.▲
   by abhgh 4y ago

   As others have mentioned on this thread, Pixel's call screener largely solves this problem. For me it's a godsend feature and prevents me from getting into a bad mood from answering multiple robo/spam-calls in a day. I seem t
7. 157.▲
   by abhgh 4y ago

   I tend to work on a lot of ML projects in my personal time, even if that means I am simply benchmarking something. Which means I generate a lot of data (in form of results) and conclusions based on them. Documentation, lavish commenting, pr
8. 158.▲
   by abhgh 4y ago

   Highly recommend his blog [1] - where he often writes bite-sized math-y articles. I have personally benefited from it; I was looking for ways to sample quickly from a mixture of Beta distributions when I stumbled upon his post on the Kumar
9. 159.▲
   by abhgh 4y ago

   It was hard for me to set aside specific hours of the day for PhD during the week - because I occasionally had early morning and late evening meetings for my day job. And on some days the work itself would spill over. So the simple policy I
10. 160.▲
    by abhgh 4y ago

    I did a part-time PhD (not in the US, so I am only responding to the latter part of your comment). It's hard, exhausting and can occasionally be frustrating. I gave up most of my weekends during the time. My sleep hours were less than
11. 161.▲
    by abhgh 4y ago

    Hmm, I guess it depends on the task and workflow. I find this easy if I have to send out the document via Gmail, WhatsApp etc. After I open up the PDF document (which I have to, anyway), the remaining steps happen on my phone. I find pickin
12. 162.▲
    by abhgh 4y ago

    Here's an even easier way to make your pdf look scanned: open it up on your laptop, take a picture of the screen with your phone using CamScanner or Adobe PDF scan. Of course this becomes cumbersome if you have more than a few pages
13. 163.▲
    by abhgh 5y ago

    I usually find articles on "foley" interesting, this included. A shout-out to the movie that introduced me to this area: [1]; this is in the "Bengali" language but still a recommended watch if subtitles are available. [1
14. 164.▲
    by abhgh 5y ago

    Agree. She had a very informative tutorial session yesterday on self-supervised learning at NeurIPS-2021. While I don't think the recording is publicly available [1], the slides are [2]. [1] https://nips.cc/virtual&#x2F
15. 165.▲
    by abhgh 5y ago

    I think of Bayes-opt as addressing multiple challenges: (a) the function being a black box, i.e. gradient info isn't available (b) tolerance to some level of noise - the author addresses this in one of his comments here (c) economic
16. 166.▲
    by abhgh 5y ago

    Maybe you meant your recommended book as related reading, but the linked text is for Bayesian optimization, which is different from Bayesian analysis. Sure the optimization uses some Bayesian ideas but the majority of the challenges are dif
17. 167.▲
    by abhgh 5y ago

    I read your article and finally got my Brother printer working on elementary OS! Thank you!
18. 168.▲
    by abhgh 5y ago

    I agree, and not just for text classification. Its a good baseline for anything that looks like text (if you train a model from scratch), e.g., URLs with some basic preprocessing. This is because FastText uses character n-grams; the consequ
19. 169.▲
    by abhgh 5y ago

    There is this anecdote in the text "Operations Research, An Introduction." by Hamdy Taha, which I consider to be a very good example of picking the right tools for the job. An OR group was hired to fix elevator wait times in an o
20. 170.▲
    by abhgh 5y ago

    Have posted about this before, but I have been working on my Python library called "compactem". A way to make traditional ML models smaller: [1]. * the core idea is to selectively use training instances to learn a compact versions
21. 171.▲
    by abhgh 6y ago

    I was scrolling down the comments to see if someone has this problem...they lost my mail too. No way to recover it then, huh? This is sad.
22. 172.▲
    by abhgh 6y ago

    Not "in lieu", but I have found SQLite to be a good companion to pandas. pandas is good when you want to lend some easily navigable structure to your data without leaving python, but becomes cumbersome (it feels cumbersome esp
23. 173.▲
    by abhgh 6y ago

    I prefer it over pandas for joins etc too. My workflow is (1) do the simple stuff in Python using pandas (2) for some of the complex stuff, I just start creating sqlite tables. If you have datasette installed, you can also view the tables (
24. 174.▲
    by abhgh 6y ago

    Well, passably. If you speak slowly and clearly. I agree about auto-correction though.
25. 175.▲
    by abhgh 6y ago

    I think this is just sign of active fads. When I wanted to pick a research area in 2007 multiple professors/students had told me to ignore databases. The understanding was RDBMSes of the day had solved all problems. There were minor ru
26. 176.▲
    by abhgh 6y ago

    This would be like the barrier function [1] for gradient based methods; which produces some challenges for those techniques. Do you know how this affects gradient free techniques? [1] https://en.m.wikipedia.org/wiki/Bar
27. 177.▲
    by abhgh 6y ago

    Would you happen to know if the speed decreases "gracefully"? As in for a SAT/SMT problem, would Coq be at least as fast as z3?
28. 178.▲
    by abhgh 6y ago

    This is a nice article. We use z3 at work (to solve client specific resource allocation problems with certain business constraints), and I personally have been interested in looking at Coq. Looking at the other comments it seems like there
29. 179.▲
    by abhgh 6y ago

    Reminds me of my homegrown version of a notifier. Essentially we had a lot of long running tasks that would emit logs. My terminal notifier: 1. Scanned for patterns in these logs periodically. 2. Send out a mail to my Gmail account with spe
30. 180.▲
    by abhgh 6y ago

    What makes this quote really interesting is that Wirth was Martin Odersky's (guy who designed Scala) PhD advisor!