karpathy

20 ms·

1. 151.▲
   by karpathy 12y ago

   Not yet. I don't know about the other papers but I really like to release code and make as much of my stuff reproducible as possible. I still have to clean a lot of it up (when you're doing research and rapidly iterating things ge
2. 152.▲
   by karpathy 12y ago

   Queries like this already work well in my experiments. For a large fraction of useful queries it's mostly a matter of engineering. My current model (which is right now at state of the art performance) doesn't yet take into account
3. 153.▲
   by karpathy 12y ago

   This Friday was the deadline for a big Computer Vision conference (CVPR), so effectively you're seeing what everyone has been working on over the last half a year :) Some authors choose to publish right away to arxiv, which is why you&
4. 154.▲
   by karpathy 12y ago

   As is usually the case with Machine Learning, its reactions are whatever the training data says they should be, or mostly undefined otherwise. For example, if there is some picture in the training data of a cartoon and it says "cartoon
5. 155.▲
   by karpathy 12y ago

   I have experiments with LSTMs as well, but I didn't include them in the paper because I couldn't get them to work significantly better than the RNN, which I consider to be a simpler model (they work comparably). That is, a simple
6. 156.▲
   by karpathy 12y ago

   Thanks! It's really not that crazy. Normally you're given an image and predict a label. For example ImageNet has 1000 labels. Here the idea is exactly the same, but we're predicting one of ~20,000 words one by one in sequence
7. 157.▲
   by karpathy 12y ago

   Extremely feasible. The model predicts a sentence for an image in ~3ms on a GPU, and ~100ms on CPU. Convolutional Networks, Recurrent Networks, in general most neural nets are very efficient at test time. The training is the somewhat expens
8. 158.▲
   by karpathy 12y ago

   Hi all, I'm the author on the Stanford article, so I can answer any more detailed/technical questions. The article doesn't mention this, but closely related work I've only very recently become aware of (if anyone is inte
9. 159.▲
   by karpathy 12y ago

   This is actually a Really good question, I should really expand on this in the guide. The simplicity of the example is deceiving you, in the sense that in a more complex circuit (than just x * y), it will no longer be the case that just sim
10. 160.▲
    by karpathy 12y ago

    Thank you! I'm a huge fan of interactive demos (see e.g. my ConvNetJS demos: http://cs.stanford.edu/people/karpathy/convnetjs/ ) because they help develop intuitions that solidify the material. And yes,
11. 161.▲
    by karpathy 12y ago

    Thanks for the upvotes! I'm a little conflicted about this being linked around too much because it is still very much work in progress. I work on this guide on a side because I think there's a lot of interest in these models out t
12. 162.▲
    by karpathy 12y ago

    The fact that RCNN does a wasteful and naive feature re-computation for every object-like region in the image is well-known in the community, and there are several papers addressing the shortcoming in very similar ways (i.e. precomputing po
13. 163.▲
    by karpathy 12y ago

    Fourier Transform became trivial to me when I noticed that it's just a basis transform, as you would do with ANY other basis. Except this basis happens to be sine and cos waves of different frequencies. i.e. you view the entire signal
14. 164.▲
    by karpathy 12y ago

    inline Show HN: or give ScholarOctopus a shot :) ( http://bit.ly/scholaroctopus ) A pet project that takes papers from these top conferences and embeds them in 2D with t-SNE based on tf-idf L2 content similarity, so you can
15. 165.▲
    by karpathy 12y ago

    Is there any parallel set of videos where they actually convey more than 10 bits of information? I'm not interested in listening to a fun synopsis of next chapter of "The story of Curiosity and the mountain". Is anyone? &quot
16. 166.▲
    by karpathy 12y ago

    Hi, I didn't notice this reply for a long time, but felt bad not to reply to something so long :) 1. Yes, good datasets exist. They grow over time and they are open. It takes a lot of time. ILSVRC is good example. There are many more:
17. 167.▲
    by karpathy 12y ago

    This also reminds me of Hemispherectomy[0] where an entire half of the brain is surgically removed in extreme cases to prevent seizures. And amazingly, especially if you do this on younger children: "Studies have found no significant l
18. 168.▲
    by karpathy 12y ago

    Google hasn't released details about their model yet, but the VGG team in close 2nd place had a very simple and beautiful "vanilla" CovnNet model, but with more careful hyperparameter settings and training/testing proto
19. 169.▲
    by karpathy 12y ago

    I think one of the main reasons is that the improvements have been drastically significant and very recent; There hasn't been enough time to convert the research code into open source code and libraries, but you can confidently expect
20. 170.▲
    by karpathy 12y ago

    I am one of the people who helped analyze the results of the mentioned ILSVRC challenge. In particular, I performed an experiment comparing Google's performance to that of a human a week ago and wrote up the results in this blog post:
21. 171.▲
    by karpathy 12y ago

    Author here. ConvNetJS is a project I maintain on a side for fun. A part of my motivation was that I wanted to make these algorithms and techniques more approachable, and easier to understand, play with and apply. One issue right now is tha
22. 172.▲
    by karpathy 12y ago

    I did similar analysis here http://karpathy.ca/myblog/?p=559 but it's a little less clean presentation. I'm quite excited about this data release- there are many interesting ML models that could be trained he
23. 173.▲
    by karpathy 12y ago

    I'm willing to bet that this is what everyone is most curious about. During the post launch news conference ( https://www.youtube.com/watch?v=8G2pRVQ1JhA ) Elon said he was pessimistic about a successful recovery of the
24. 174.▲
    by karpathy 13y ago

    I did a similar analysis to the one posted here and computed a similar heat map to the one you describe, but I marked a submission as successful when it went from new -> front page, not when it hit 100 points. The result is in ~middle of
25. 175.▲
    by karpathy 13y ago

    I remember that they had an impressive demo at the NIPS Deep Learning Workshop in December (probably a highlight for me and a few people I talked to afterwards) of a reinforcement learning agent playing Atari 2600 Games where the input cons
26. 176.▲
    by karpathy 13y ago

    By the way, just this evening I formatted the accepted papers in a nicer way: (okay, I'm biased) http://cs.stanford.edu/people/karpathy/nips2013/ The page allows you to toggle LDA topics and off to brows
27. 177.▲
    by karpathy 13y ago

    I'm part of this "renewed interested in chess", as I've played competitively when I was in high school but then stopped for several years until I came across this year's championship a few weeks ago. I've tried
28. 178.▲
    by karpathy 13y ago

    Interesting, Based on my own discussions with friends and my own experience, it's considered strictly annoying and done likely to save bandwidth or something. Can you explain what you like about it?
29. 179.▲
    by karpathy 13y ago

    The official commentary and analysis, imo, was not very good though. I was watching this feed live to see the players, but for audio I was over at http://www.twitch.tv/chessnetwork . Jerry from chessnetwork did a wonderful
30. 180.▲
    by karpathy 13y ago

    Content aside, I love the idea of visualizing various skills and their relationships in this World of Warcraft - like style! Nicely done. Khan Academy could be taking notes and inspiration to replace their current Knowledge Map: https:&#x2