karpathy

13 ms·

91.
by karpathy 9y ago
To add to this, there are many not-so-great side effects due to all of the complexity even if the code is all correct and perfectly reproduced. For instance, I suspect a large fraction of claimed "better than baseline" results in
92.
by karpathy 9y ago
+1 that the meta-learning spin on this approach is the really interesting part. The normal approach would be as follows: "you want to stack 6 blocks on one another? great, let me collect 1,000 examples of doing that in VR, and I'l
93.
by karpathy 10y ago
I wasn't too surprised. If you look at the Vector Institute website ( http://vectorinstitute.ai/#people ), you'll see many additional well-known researchers involved, in addition to Geoff Hinton: David Duvenaud, San
94.
by karpathy 10y ago
The thing I always found perplexing about the idea of an ancestral simulation is - how do you initialize that? E.g. suppose we are in the year 3,000 and we have super advanced computers and I'd like to simulate the year 2017. How do I
95.
by karpathy 10y ago
Not sure if I understand "automatically generating labels to a SL network". I don't believe RL is used in this setting. RL is about learning expected-reward-maximizing policies for environments that you get to interact with.
96.
by karpathy 10y ago
If it makes you feel any better, I've been doing this for a while and it took me last ~6 weeks to get a from-scratch policy gradients implementation to work 50% of the time on a bunch of RL problems. And I also have a GPU cluster avail
97.
by karpathy 10y ago
Agreed with OP on HoloPoint and other apps (e.g. AudioShield is another favorite of mine) being a surprisingly intense workout. I sweat quite a bit with these games after ~15min, and this is definitely a problem. On Vive, the soft padding a
98.
by karpathy 10y ago
Another interesting aspect is that AlphaGo gets a larger advantage over human players if it plays novel moves and forces rare situations that the human players are not familiar with. So if AlphaGo is allowed to self-play a long time and &qu
99.
by karpathy 10y ago
Tom White has done a lot of fun work recently in this direction https://twitter.com/dribnet . You can absolutely do it with faces.
100.
by karpathy 10y ago
It seems that a larger-than-I-expected fraction of the response to my post concentrates on a very small part of it, especially the part where I enumerate some considerations for thinking about whether a PhD might be a good fit for you. By f
101.
by karpathy 10y ago
Google Drawings (or Slides) is good, and Pages/Keynote is very nice too.
102.
by karpathy 10y ago
I happen to :) http://cs.stanford.edu/people/karpathy/main.pdf
103.
by karpathy 10y ago
Thanks! The option to contribute additional links would be a great feature. As to discussions about papers there are plans (semi-related to arxiv-sanity) in motion to do that well and correctly, not just from me alone. I think we'll se
104.
by karpathy 10y ago
Thanks! Email alerts are one of top requested features, it's definitely on my shortlist for the next feature to incorporate. Another feature I'd like to add is an ability to follow people, but I'm worried about the exact impl
105.
by karpathy 10y ago
I wrote http://www.arxiv-sanity.com/ (code is open source on github: https://github.com/karpathy/arxiv-sanity-preserver ) as a side project intended to mitigate the problem of finding newest relevant wo
106.
by karpathy 10y ago
A human is very good at one-shot learning but CNNs are actually not too terrible either (and this is also an active area of research, e.g. see http://www.arxiv-sanity.com/1603.05106v2 ). A human might take advantage of good
107.
by karpathy 10y ago
People rarely train on billions of images, we're usually around the scale of ~million. This already works quite well in many respects. A back of the envelope calculation assuming about 10fps vision gives ~1B images by age of 5. And hum
108.
by karpathy 10y ago
I don't select anything by hand, it's listed by date, as it comes from arxiv API. The 2 day lags are due to weekends, when arxiv does not update.
109.
by karpathy 10y ago
I developed and maintain Arxiv Sanity Preserver ( http://www.arxiv-sanity.com/ ), one of the Arxiv overlays the article mentions. I built it to try address some of the pains that the "raw" arXiv introduces, such as
110.
by karpathy 10y ago
Sorry about that! There's a lot to cover for one blog post to do satisfyingly. I encourage you to check CS231n for a more thorough treatment where we also discuss, for example, the tradeoffs of different activation functions like tanh(
111.
by karpathy 11y ago
this is very nice! I think that the reason swiss roll doesn't work as easily might be because of initialization. In 2 dimensions you have to be very careful with initializing the weights or biases because small networks get more easily
112.
by karpathy 11y ago
Might be easiest to think of your privacy as a currency similar to $. The real cost is 800. They can offer you $600 and 200 in privacy, or they could offer you $800, which people would throw a fit over and not buy. This is also not entirely
113.
by karpathy 11y ago
I thought about this quite a bit. I don't think I want downvoting. And I don't want effectless upvoting. The way it is right now is that you can add a paper to your library, and it will then feed into your personal SVM as a positi
114.
by karpathy 11y ago
the right tool for the job :) In this case I'm perfectly happy with SVMs over tfidf bigrams and where that places you in the tradeoffs space.
115.
by karpathy 11y ago
Yeah sorry about that - I see cryptic errors in the server logs popping up at random. I think it must be something to do with the scale of requests coming in and breaking the site in some way I don't currently understand. I have near z
116.
by karpathy 11y ago
Imagine waking up every morning with 50 new arxiv papers uploaded that night. You panic and quickly scan through the papers - any of them could be very related to your research, or scoop your latest idea, or have good ideas you can use in y
117.
by karpathy 11y ago
Is there the scramble/analysis of the 4.90s solve somewhere? Looks to me like a nice fast cross, standard F2L, standard OLL, and then a clean PLL skip? edit: found something on page 16 of a thread: https://www.speedsolving.c
118.
by karpathy 11y ago
apparently it's a "terrible hack" with text files https://twitter.com/kcimc/status/668111582296190976
119.
by karpathy 11y ago
So here's what's funny about the leaderboard. The NeuralTalk submission (near bottom of the list) is in fact a re-implementation of the Show and Tell model (my paper and Oriol's are basically identical models - we plug in the
120.
by karpathy 11y ago
In image captioning specifically we are in a dire need of data. To give you an idea, MS COCO is ~100K images. ImageNet Challenge is 1.2 million. The dataset is quite miniscule, which restricts the richness of models we can explore, and forc

More ›