alexcnwy

24 ms·

1. 271.▲
   by alexcnwy 9y ago

   I think there can be a match but it's a spurious match.
2. 272.▲
   by alexcnwy 9y ago

   I'm also curious if this makes sense?
3. 273.▲
   by alexcnwy 9y ago

   You can re-pose it as a regression problem to detect the coordinates of the 4 corners in the case of a single object in the image.
4. 274.▲
   by alexcnwy 9y ago

   Also, see things like the DeVISE paper that replace the 1-hot encoding output on the 1000 imagenet paper with word vectors so the model can generalize beyond 1000 classes: http://citeseerx.ist.psu.edu/viewdoc/download?d
5. 275.▲
   by alexcnwy 9y ago

   BTT also lets you customize the touchbar. And go full Minority Report with gestures on the trackpad.
6. 276.▲
   by alexcnwy 9y ago

   I run a Machine Learning consulting business. Been fully remote without a 'home base' for about a year. Really loving it so far and plan on continuing for the foreseeable future! My advice would be to spend at least a month in eac
7. 277.▲
   by alexcnwy 9y ago

   why the downvotes?
8. 278.▲
   by alexcnwy 9y ago

   Why no South Africa?! :)
9. 279.▲
   by alexcnwy 9y ago

   This brings up the interesting question of whether it's ok to publish the trained model and not the data? I'd imagine it's sufficiently derivative as to be allowed? What if you wrote a model that produced summaries of the boo
10. 280.▲
    by alexcnwy 9y ago

    If your clients are in Europe then being 5 hours ahead is actually better than being in the same timezone (especially if you're nocturnal and only wake up at 11am ;)
11. 281.▲
    by alexcnwy 9y ago

    Do data augmentation for the geographic question - overlay the lat-lons with other spatial data (e.g. demographics, distance to nearest mall, etc.)
12. 282.▲
    by alexcnwy 9y ago

    I find it curious that machine learning is a branch of computer science given how much of it is actually statistical learning theory. The go-to graduate level machine learning text "The Elements of Statistical Learning" was writte
13. 283.▲
    by alexcnwy 9y ago

    Great product and congratz on your success so far! Would you mind please elaborating a little more on how you're thinking about theme-specific sentiment (for the NLP students here :) - is it along the lines of Socher's aspect-spec
14. 284.▲
    by alexcnwy 9y ago

    It's great to have this all in one place - the change in notation across papers can be quite confusing! I just wish there was something like this with code examples. It's so frustrating when papers say things like 'we omit bi
15. 285.▲
    by alexcnwy 9y ago

    Firstly, while I think it's beneficial to learn multiple languages (python, R, matlab, julia), I'd suggest picking one to avoid overwhelming yourself and freaking out. I'd suggest python because there are great tools and lots
16. 286.▲
    by alexcnwy 9y ago

    Humans can't do _everything_ better than insects, does that mean insects are more intelligent than humans? Chimpanzees have better short term memory on certain tasks than humans [1] - humans not being better at _everything_ than chimpa
17. 287.▲
    by alexcnwy 9y ago

    Honestly I think you're playing a language game. I'm sure you're familiar with the line of reasoning that if you asked someone 50 years ago to describe tasks that require intelligence, they'd for sure say recognizing obj
18. 288.▲
    by alexcnwy 9y ago

    Actually Andrej Karpathy showed human level error on imagenet is around 5% while the winning model from 2017 had an error rate of just 2.25%.
19. 289.▲
    by alexcnwy 9y ago

    There was a great podcast with Soumith Chintala on the O'Reilly data show a couple of days back with more info on PyTorch and how it differs from Theano and Tensorflow: https://www.oreilly.com/ideas/why-ai-and-mach
20. 290.▲
    by alexcnwy 9y ago

    I believe Keras is being made into an open spec that PyTorch could be made to run with.
21. 291.▲
    by alexcnwy 9y ago

    You can definitely get by just using AWS and the flexibility is great but it can get expensive. The Tiramisu model from lesson 14 takes like 25 hours to train and at $0.90 per hour, it adds up pretty fast...
22. 292.▲
    by alexcnwy 9y ago

    If your data is in S3 then data transfer fees can suck but def cheaper than $ 0.90 per hour if you're regularly leaving models to train overnight :)
23. 293.▲
    by alexcnwy 9y ago

    Honestly can't recommend this course highly enough. It's definitely not perfect - the notebooks are not commented and the material does tend to jump around a bit - but what it does do, it does extremely well. This course will teac
24. 294.▲
    by alexcnwy 9y ago

    Totally agree The work on textual understanding is definitely still early days (though the press release makes it seem like it's understanding the entire LOTR, it's just reading a very structured short version) but already crazy i
25. 295.▲
    by alexcnwy 9y ago

    Can just screencap the rendered page and use OCR. If someone wants to scrape your website badly enough, they'll find a way.
26. 296.▲
    by alexcnwy 9y ago

    Plus very easy to parallelize
27. 297.▲
    by alexcnwy 9y ago

    I'm not sure if it's officially out yet but if you search YouTube, you will find it ;) https://m.youtube.com/watch?v=6lTyqrrWVQ0
28. 298.▲
    by alexcnwy 9y ago

    I'd suggest lecture 14 of the fast.ai mooc on some advice on feature engineering for timeseries data to make it possible to model with something like regression (or a non-recurrent neural net - can work surprisingly well)
29. 299.▲
    by alexcnwy 9y ago

    You have to request access to on demand P* instances so I suspect some people can't get as much allocation as they need and end up turning to the spot market. Makes sense in constrained supply regime.
30. 300.▲
    by alexcnwy 9y ago

    Data transfer between EC2 and S3 in the same region is free but incurs costs between regions.