argonaut

26 ms·

241.
by argonaut 9y ago
No, no, no. > whatever penalty Keras might impose is equal in the two cases. The penalty Keras imposes when using Tensorflow depends on its Tensorflow implementation. The penalty Keras imposes when using MXNet depends on its MXNet implem
242.
by argonaut 9y ago
The PlaidML benchmarks are suspect. They compare to Keras + Tensorflow, which is a really unfair comparison since 1) Tensorflow is probably the slowest of the big deep learning frameworks out there (compared to PyTorch, MXNet, etc.), and 2)
243.
by argonaut 9y ago
> shallow neural networks can be universal function approximators in a trivial sense because they can be lookup tables, but they are really not expressive enough to generalize well You've got it flipped. If anything, shallow neural
244.
by argonaut 9y ago
The TFLOPS numbers are not directly comparable. The TPUs use reduced precision in some areas, whereas I am guessing the Titan V numbers are based on single precision operations.
245.
by argonaut 9y ago
Yes, but that point is probably when you have hundreds of engineers, at which point you can afford to have (and probably will already have) a few engineers working on internal tooling.
246.
by argonaut 9y ago
For the first concern you have things totally flipped. A monorepo actually seems best suited for a small company, with a small codebase and a small set of services. If anything, the stage of company where it makes most sense to have many sm
247.
by argonaut 9y ago
This is a meme that keeps getting repeated, and I don't know why. Tensorflow, for example, despite several years of development, does basically little to no graph optimizations and for tons of tasks ends up much slower than PyTorch &#x
248.
by argonaut 9y ago
This is one of those catchy but actually unverified memes that gets endlessly propagated. "Nevertheless, I believe that a world-champion-level Go machine can be built within 10 years" - Feng-Hsiung Hsu, 2007 (researcher who work
249.
by argonaut 9y ago
The editorial department of the WSJ is almost certainly "compromised" (in the sense that they have political opinions and loyalties), but that is true of even the NY Times. The actual reporting at the WSJ is still legit.
250.
by argonaut 9y ago
It seems obvious now, but you only have to look at an HN thread 6 years ago discussing Yuri Milner to see that many people in tech (or at least on HN), including pg, did not believe Yuri was an extension of the Russian state: https:/&
251.
by argonaut 9y ago
The critic idea seems interesting but doesn't really get to the question of whether/to what degree the GAN is just interpolating the training data. It seems more of a useful tool for diagnosing GANs.
252.
by argonaut 9y ago
Worth pointing out the main criticism of GANs, which is that right now researchers don't really have a way to tell if a GAN is just copying and pasting the training data or not (there is no "test set" unlike in supervised lea
253.
by argonaut 9y ago
No, it's not widely used anymore, although I wouldn't call it rare either. The most popular image CNNs don't use them: ResNets, Inception, VGG, fully convolutional nets, etc. The Google query gwern cites is highly mislead
254.
by argonaut 9y ago
Your example actually runs counter to the idea that we're seeing a massive breakthrough. Deep nets on MNIST for recognizing numbers were done 20 years ago.
255.
by argonaut 9y ago
It's funny that you mention Relu. People have recently trained Imagenet networks using sigmoid/tanh (e.g. the activations that were used decades ago) on GPUs and they train just fine. They train a bit slower is all. Not the breakt
256.
by argonaut 9y ago
I thought this analysis article was really thought-provoking and now I have uncertain and mixed feelings about the referendum: https://www.nytimes.com/2017/09/29/world/europe/independence... . Basica
257.
by argonaut 9y ago
Not really, no. Saying we are at year five of Deep RL is about as informative as saying we are at year five of deep learning. Reinforcement learning as a field goes back decades.
258.
by argonaut 9y ago
> that you don't raise money if you don't need it or don't anticipate you will need it? Slack anticipates they might need the money in the future.
259.
by argonaut 9y ago
Transformer is heavily reliant on supervised learning. Not sure where you're getting that from.
260.
by argonaut 9y ago
But in many cases using credit will save you money. Two examples would be credit cards with their rewards (as long as you never pay interest) and house/car loans if you manage to get a low rate and invest what you don't pay up f
261.
by argonaut 9y ago
No it didn't. I've been actively reading/commenting on HN for 5 years. Bitcoin got a lot of hype 3-4 years ago. The only difference is it gets massive amounts today.
262.
by argonaut 9y ago
Warning: Uber counts the entire Uber Pool fare are revenue, which throws off these numbers a lot since I would guess that Uber Pool is the most popular option where available. Source: https://www.bloomberg.com/news/ar
263.
by argonaut 9y ago
People (rightfully) pointing out that Siri sucks compared to competitors should realize that Apple is making a very serious tradeoff in favor of privacy. It's a huge handicap when you're only allowed to train models on device usin
264.
by argonaut 9y ago
You really can't compare companies in different industries. Most Google employees are "knowledge" workers with some barrier to replacement. Most Walmart employees are performing mundane service work that is largely replaceabl
265.
by argonaut 9y ago
I'm not an expert either, but what if the object-size pair is not in your database of known object-size-distances.
266.
by argonaut 9y ago
It's really a question of timeframe. You are completely right that we theoretically only need two cameras. But by limiting yourself to 2 cameras (or cameras + other non-lidar sensors) you're just making the problem a lot harder
267.
by argonaut 9y ago
If a company's solution to self driving cars with just two cameras requires developing a machine learning "model of the world" (I don't think it does, but it does make it a much harder research problem), then they are go
268.
by argonaut 9y ago
To my limited knowledge, they don't. Deep nets are mostly used by the self-driving companies for processing camera data for image detection and segmentation tasks like detecting other vehicles, recognizing street signs, etc. Presumably
269.
Only 802 People Told the IRS About Bitcoin–Lawsuit(fortune.com)
4 points by argonaut 9y ago|0 comments
270.
by argonaut 9y ago
This is such a backwards-looking point of view. 0. The fact is, I want a algo timeline. There are far too many posts in a day for me to read everything. 1. People made the same complaints about Facebook's algorithmic feed. Very few u

More ›