emcq

31 ms·

1. 271.▲
   by emcq 11y ago

   Good point - it is kind of a lazy shortcut. Next time I'll type out the extra words :) It probably would have been sufficient just with the approx ~
2. 272.▲
   by emcq 11y ago

   It's easy to think of reasons why this isn't a sound investment. However, here are some thoughts why this may be sensible: * This is largely Google's investment to cover possible future success of Facebook's Occulus or M
3. 273.▲
   by emcq 11y ago

   A little confused about the authors decision to do operator overloading with the pipe operator rather than a function chaining interface for the functional bit. Wouldn't a function chaining interface be simpler and use less magic?
4. 274.▲
   by emcq 11y ago

   You will probably be interested in Kolmogrov complexity, which studies this using an information theoretic lens. However it is just an abstract concept and not really something you can measure precisely.
5. 275.▲
   by emcq 11y ago

   Reasons like these are why I get frustrated when people claim the only reason why NNs are popular now is because of computational advances. This isn't true - there is lots of little learnings that have been developed in recent time to
6. 276.▲
   by emcq 11y ago

   I don't think so - it is fairly common in simulations and statistical analysis as a way to explore the robustness of your results given small changes within the tolerances you expect for your model. Usually it is applied to the input
7. 277.▲
   by emcq 11y ago

   I've tried to do a test like this and the results surprised me. While the testing conditions weren't perfect, I tracked how much time it took to port Fortran code to C++, and how much time it took to write the C++/CUDA optimi
8. 278.▲
   by emcq 11y ago

   Completely agree. I like Bret's stuff and love Julia but that section seemed very weird and almost preachy. He seems a little out of his element because a lot of physicists doing earth modeling use compiled languages like Fortran and C
9. 279.▲
   by emcq 11y ago

   Dynamic queries and data serialization. Imagine sending just a text query (in SQL) around to a sensor network. It would be a fair amount of work to reproduce that on your own but seems like it can open up some nice features for distributed
10. 280.▲
    by emcq 11y ago

    Blast from the past! If you think this kind of less traditional circuit design is cool, you should also check out some of the more recent stuff in low-power asynchronous architectures: http://vlsi.cornell.edu/research.php
11. 281.▲
    by emcq 11y ago

    I get excited whenever we can improve performance with manipulations to the training sets such as data transformation tricks (inducing variances like rotations, translations, whitening [1] etc), labeling tricks (such as those like LeCun or
12. 282.▲
    by emcq 11y ago

    Sure, I agree the seatbelt example comes from not having enough labeled classes. Feels a little bit more like the realm of semi-supervised learning, but I'd love to learn about interesting approaches in that case too :)
13. 283.▲
    by emcq 11y ago

    Are there any interesting approaches you've seen to help reduce overfitting with these class imbalance issues?
14. 284.▲
    by emcq 11y ago

    After training, you could likely do this on a single node. You don't need that much compute for speech to text and then information retrieval.
15. 285.▲
    by emcq 11y ago

    I'm not sure about their numbers, but with a little trick it can be estimated as 1-e^(-2^(n-x-1)), where x is the number of bits and n is log_2 of the number of random draws. When n and x are relatively close, you should be able to cal
16. 286.▲
    by emcq 11y ago

    Agreed, but it can still be a useful technique if you ever need to sort your values by time, in which case it's logically a tuple masquerading as a physical single id.
17. 287.▲
    by emcq 11y ago

    While not an exact answer, you can certainly replicate the behavior of complex neurons as a composite of several simpler ones. Check out fig 9 in http://www.research.ibm.com/software/IBMResearch/multimedia/...
18. 288.▲
    by emcq 11y ago

    I actually have "research software engineer" on my resume and you're right, it doesn't do much. In fact I don't think this is uncommon, at least in the US. Myself and former colleagues indeed have no trouble working
19. 289.▲
    by emcq 11y ago

    Is this surprising? I think it's expected, when most published numbers from Google suggest that this is only 2x over DistBelief, which Adam Coates improved upon 6x ( http://www.cs.stanford.edu/~acoates/papers/C