yablak

14 ms·

181.
by yablak 9y ago
This. The article basically reverses its own argument with this statement. The real purpose of Facebook's license is to stop any sufficiently large company (i.e., any company with the wherewithal to sue Facebook) from using their tech
182.
by yablak 9y ago
If you're using TF 1.2.1 or above, you want to replace this line: cell = tf.contrib.rnn.MultiRNNCell([drop_cell] * num_layers) with: cell = tf.contrib.rnn.MultiRNNCell([create_cell() for _ in range(num_layers)]) otherwise you'll b
183.
by yablak 9y ago
As a tensorflow dev, with several years experience writing clojure at a previous startup, I believe this is the most natural language for writing tensorflow models (or theano. basically any graph building dsl that has any real complexity an
184.
by yablak 9y ago
You're entering an active and mainly successful research field: https://nips.cc/Conferences/2016/Schedule?showEvent=6213 TL;DR: if a human eye can differentiate, GANs can learn to differentiate given enough t
185.
by yablak 9y ago
A parser combinator may be more powerful? Not sure. The python parsec package is really terribly documented, but there's a good example here: http://www.valuedlessons.com/2008/02/pysec-monadic-combinato... E
186.
by yablak 9y ago
Ticket reopened.
187.
by yablak 10y ago
(nowadays it's usually the other way around, of course ;)
188.
by yablak 10y ago
Before computers had the kind of processing power they do now, engineers would build analog circuits to solve differential equations. A modern-day quick link: http://www.dummies.com/education/science/science-electr
189.
by yablak 10y ago
The first mistake is using shared_ptr to begin with. It's often unnecessary (as the article says, one can often get away with using unique_ptr). When a unique_ptr is not sufficient, a shared_ptr often obfuscates the true owner of the
190.
by yablak 11y ago
No. Most of layers in a CNN perform convolutions with kernels. This is not the same as standard DNNs that do a full matrix multiply. Convolutional kernels allow you to use many fewer variables to perform the forward layer operation; and C
191.
by yablak 11y ago
Deep convolutional networks have enjoyed great success in image recognition and understanding in the last 3+ years (see the ImageNet competition winners). One of the main reasons people believe they are so powerful is due to translation, r
192.
by yablak 11y ago
You're right. There is not currently a theano.scan equivalent that dynamically loops over a dimension of a tensor. That said, you can do a lot with truncated BPTT and LSTM. See the sequence modeling tutorial on tensorflow.org for mor
193.
by yablak 11y ago
Can someone explain to me how Relay compares to gRPC?
194.
by yablak 11y ago
"The internal management application of this is that it’s always a bad idea to set one’s employees too much against one another. You want to find ways to differentiate people’s roles. Frame it this way: If you were a sociopathic boss w
195.
by yablak 11y ago
BTW, I believe the "generalization" of the taylor expansion mentioned here is the Laurent series ( http://en.wikipedia.org/wiki/Laurent_series ).
196.
by yablak 11y ago
I remember rediscovering fractional derivatives in college after learning about the Laplace transform derivative formula. I even called them fractional derivatives; and have a maple notebook somewhere in which I identified some fractional
197.
by yablak 11y ago
fractional derivatives have found some uses in signal processing and control systems. you can search that exact term in google scholar.
198.
by yablak 11y ago
Probably not. gamma(N) is the single mostly-analytic function for which gamma(x+1) = x*gamma(x). This is a requirement.
199.
by yablak 11y ago
From the functional form of the author's definition, it looks like yes. However, to connect the two definitions probably requires using Cauchy's integral formula ( http://en.wikipedia.org/wiki/Cauchy%27s_integ
200.
by yablak 12y ago
They wrote Torch7...
201.
by yablak 12y ago
This is generally taken care of using checksums, assuming you trust the source. Of course, trusting the source is the hard part. CyanogenMod, for example, has a chain of trust for their releases; though they host their own binaries.
202.
by yablak 12y ago
Climate didn't have this service available until recently. Anyway, they're probably using the same rainfall data (URD and/or AHPS)
203.
Investment Riches Built on Subprime Auto Loans to Poor(dealbook.nytimes.com)
62 points by yablak 12y ago|56 comments