leecho0

The article claims that RL is simplistic because it uses an unreasonable amount of data. However, recent advances are significant because it uses unreasonable amount of data. As an example, I don't expect to be as good as Michael Jordan no matter how much I play basketball, or beat Garry Kasparov no matter how much I play chess. There's a fundamental flaw to my learning algorithm that prevents me from becoming good at something even if I have infinite experience.

Recent RL research about Policy Gradients / On Policy vs Off Policy / Function approximation / Model-based vs model-free are all research about how to get good at something with a lot of practice. RL has been around for a long time, discussions about higher level learning / planning has been done over and over. One doesn't discount the other. One deals with how to structure the learning problem that you can continue to get better with more experience (RL problem), while the other is about how to use higher level learning to speed it up.

The autoencoder converts an image to a reduced code then back to the original image. The idea is similar to lossy compression, but it's geared specifically for the dataset that it's trained on.

According to the defaults in the code, it uses float32 arrays of the following sizes:

  image: 144 x 256 x 3 = 110,592 
  code:  200

Note that the sequence of codes that the movie is converted to could possibly be further compressed.

I haven't read the paper, but according to the article, "The treatment, known as immunotherapy, uses the body's immune system to attack cancerous cells." Rather than trying to kill the cancerous cells, the drugs allow your body to attack the tumors.

To give some background, your own body already tries to kill cancerous cells through cytotoxic t cells. The tumors that get serious are the ones that escape the immune system. However, strengthening the immune system has its own problems, since it can attack your own cells (autoimmune diseases), or cause other problems (like Crohn's disease).

So I'd imagine the cancer cells probably won't get immunity to the drugs, but other serious side effects could come from the treatment.

Which is why Line could really make it that big. Line competed and won against Whatsapp, Facebook messenger, Skype, etc for iOS and Android in many Asian countries (not just Japan and Korea).

* It's a good product -- people like using it, and they keep using it.

* They know mobile marketing very well -- it's common to see Line and affiliated apps dominating the charts in the appstore.

* They know how to be international -- they've successfully marketed across multiple cultures.

* They are very innovative -- they constantly try out new business models and release multiple features.

It's no wonder that so many companies are copying them -- Facebook, Path, Wechat. It's a company that's worth paying attention to.

(edited for format)

good points,

#2 may be difficult for certain uses, since it's the system that displays the pop-up notification for push notification. But I agree it should be removed if possible.

#6 can be done with how multitouch is handled, like the wrist guard feature in note taking apps like penultimate. In most cases, just allow multitouch interactions with objects on screen.

#7, question for parents: what's the monetization policy that would work for you? - one free app from the developer, all others cost money - free app with in-app purchases, tucked away somewhere for parents - no "sample" app, all apps cost money

are they open to visitors?

Any thought about what to do if you didn't follow the advice? Our startup is 10 months in trying to convince a non-early adopter to buy our enterprise product. They want the product, but like the article claims, they are too risk adverse to just buy it without months of testing.

Should we stick it through to make it as main stream as fast as possible (we chose the client for their effectiveness as a reference) or backpedal and look for early adopters?

I'm curious, how much faster would it run if these tests were written in native code?

bonus tip:

don't forget to add:

  # -*- coding: utf-8 -*-

also, if you're using vim, make sure your encoding as well as your fileencoding are correct (they're different):

  set encoding=utf-8
  set fileencoding=utf-8

Wealthy foreigners already do buy their way into the States through the EB-5 visa.

http://en.wikipedia.org/wiki/EB-5_visa

The startup visa seeks to change the EB-5 visa to include people who aren't wealthy, but would also create jobs in the US by starting viable companies.

So making a mobile OS and having image recognition technology makes Google likely to be at the center of the robotics industry?

It seems like a stretch.

I'm quite confused by the technology behind this. I tried searching for "leather wallet," and it said that the term is being established, and currently has no results (while wallet gave me quite a few). Does this mean that you only index pictures after someone typed in the term?

I like the front end, but it seems like the search engine needs some work to find useful pictures.

So the rats were more stressed out from cold water than getting killed and getting their brains examined? interesting.

I think what you meant was that there is a huge gap between spoken languages and programming languages, feature wise.

To make the argument more convincing, don't use English. The grammar is ambiguous. Lojban, a created language, is not. It is being used as a spoken language (though it's not very popular yet). Still, it is being used very differently from a computer program in a fundamental way.

One of the major issues is the concept of object definitions. Object-oriented languages have a very specific meaning for what an "object" is, while spoken languages do not. Rather than creating a specification for what data and functions an object contains, spoken languages group objects by properties. Ambiguity aside, it's a problem of whether you're defining it top-down, or bottom-up. Prototypes are closer to the way we use language for object definitions, but it's still not quite the same.

Another issue is how the language is used. A random portion of the source code is completely non-sensical to a computer, while a random paragraph of a book can still be meaningful. The portion of the source code could make sense to a programmer, maybe enough to reconstruct enough of it, but no computer can process it.

So, to go back to the original point, programming languages and spoken languages are both supposed to convey ideas. Spoken languages do a much better job, hands down. OP used length as a rough measure, but the idea is valid. The question is how can we make computers understand spoken languages in a meaningful way. Answer that, and you'll have a better programming language.

However, once programming languages reach a certain maturity, the marginal benefits of a better language will be dwarfed by the effort it would require to utilize it.

but it doesn't do a good job of explaining what to do when your database isn't up to date. Geo django also gives horrible error statements (none at all when the libraries aren't installed right).

replace threshold with parameter,

classifiers may have different parameters other than threshold

Area under ROC curves (AUC) are used to evaluate how good the model is considering many parameters can be chosen for it.

However, when you actually want to make a prediction, you still need to pick a value for the parameter (a point on the ROC curve). So F-measure is often more useful for evaluating the predictions. While AUC are more useful at evaluating classifiers.

I think I might just have to try it out at home sometime.

Anyone know what would make a good salt bridge that's semi permeable to na/cl (or how to treat the polystyrene)? Would an U shaped jello mixed with acid/base work?

Great idea! Instead of trying to fight osmosis, use it. It's one of those "why didn't I think of that" inventions.

I was wondering more about like regression of customer data, creating recommendation engines, or something along those lines.

I'm curious what you guys use ML for

Question, if a restaurant gets on it once, will it ever be able to get on again?

If there's a "buzz" at a restaurant, there will be a lot of people there, and it will be harder to find a seat. Of course, it'll be a good sign if you have this problem, but it may be something to watch out for.

Good points. Genetic testing is still at its infancy, so we really don't know that much about the relationship between genetic code and actual diseases. Very expensive large scale testing is needed to improve accuracy, so the best alternative is to tell the user how they came up with the numbers. I think these companies do know that their techniques aren't very useful right now, but they want to get a head start on what will inevitably be a big part of our society.

FTA: it’s that no matter your genetic makeup, most common diseases can be prevented with proper diet and exercise.

Some nit-picking points: "myth1 - having your genome read is like hex-dumping the ROM of your computer" Many many interesting high throughput sequencing methods are coming out, and they would actually give you the full sequence of your genome. However, to follow your analogy of hex-dumping, we don't know the assembly language to decode what the genome does. What these companies do give you, though, are the places in the genome that are related to diseases, which would be enough if we had enough data about the relationship between the disease and the genetic code.

What's more interesting, is reading mRNA with those sequencing or microarray technologies, which tells you what code is currently being run in the cell. That would not only tell you your inherited disposition for disease, but the current status of your health -- strong correlations were found for certain genes and cancer fatality. But of course, it's not available to consumers yet.

"The majority of SNP mutation-based “predictions” are correlative, not causative" You need to do very specific tests to find out causation. Correlation don't mean causation, but they are significantly related, and it makes sense to use it as a predictive tool under certain circumstances.